Changelog

All notable changes to the high machines website (07-website) are documented in this file. Newest version first.

The format is inspired by Keep a Changelog — and this project adheres to Semantic Versioning. Two JP-specific additions: a per-version 🔭 Vision TL;DR (so each version is scannable in one breath) and an ⚠️ Known Issues section (so partial ships and in-flight work are surfaced, not hidden).

The pre-v0.25.0 history is preserved in the 📚 Legacy entries section below in its original prose form. A copy-paste-ready 📋 Template lives at the very bottom of the file.

For the active user-story backlog, see BACKLOG.md. For the design doc, see SPEC.md. For the spec-revision history (v0.1 → v0.12), see SPEC.md §Changelog — that tracks changes to the design doc, not the app.

[Unreleased]

🔭 Vision

House-cleaning sweep (lesson-#0a + #20 follow-up to US60 + US31 closure). The 957c958 US31 closure commit message claimed "Quick Stats 1 in progress → 0 in progress" but the actual diff left the line stale — this commit backfills the Quick Stats flip + Next-sprint US31 tail + Story-map US31 indicator + drops the giant Previous-version Status chain in README.md line 5 + SPEC.md line 3 (US60 said drop but US58/US59 re-added it post-US60 commit) + flips 6 stale [ ] task checkboxes for US52/53/54/56 that said "deferred to JP's machine" but the container is live on v0.44.0 since the post-US58 rebuild (~2026-06-29). No version bump — content-only.

📝 Documentation (house-cleaning sweep, 2026-07-07)

• BACKLOG.md Quick Stats pipeline bullet — flipped + **1 in progress** (US31 … 350 MB target remaining) → + **0 in progress** (US31 closed 2026-06-30 as content-only revision in commit 957c958 — original multi-stage ship landed as v0.21.0 on 2026-06-10, 350 MB follow-up target explicitly abandoned per JP's inspection). Pipeline total still 62 US; 56 shipped / 0 in planning / 0 in progress / 4 future sprint.
• BACKLOG.md Next-sprint (5+) line — US31 trailing entry rewritten from **in progress** (multi-stage Dockerfile shipped, 350 MB target remaining with pnpm deploy) to **closed 2026-06-30 as content-only revision** (commit 957c958; original multi-stage ship landed as v0.21.0 on 2026-06-10, 350 MB follow-up target explicitly abandoned per JP's inspection — marginal 513 MB savings on a single cached image doesn't justify the architectural cost).
• BACKLOG.md Story map (dependencies) — US31 arrow flipped from 🚧 partial ship (680 MB) to ✅ closed 2026-06-30 (content-only, 863 MB, no version bump — original multi-stage ship landed as v0.21.0 on 2026-06-10, 350 MB follow-up target abandoned). (The (680 MB) was a stale early-partial-ship measurement; canonical at Quick Stats line 448 is 863 MB per the 2026-06-26 measurement.)
• BACKLOG.md Last Updated line — bumped from 2026-06-30 14:30 UTC (when US60 closed) to 2026-07-07 08:30 UTC with the full housekeeping-sweep summary. Documents the lesson-#0a violation in 957c958 (commit message claimed Quick Stats flip, actual diff did not land it) so the next audit doesn't repeat the same check.
• BACKLOG.md deferred-JP checkbox sweep — 6 stale [ ] task checkboxes flipped to [x] with **Done** (3) or **superseded** (3) annotations, since the prod container is live at https://localhost:8001/ on v0.44.0 (verified 2026-07-07) and the "deferred to JP's machine" condition has been met by every post-US52 rebuild cycle. Affected: US52 T52.20 + the pnpm build+Container live acceptance bullets; US53 T53.13; US54 T54.10/11/12 + the DoD pnpm build+Container live+README+SPEC+... bullets.
• README.md line 5 — Status header rewritten: now leads with **Status:** v0.44.0 (2026-06-30, US58 ship — investor-facing /deck route...) and replaces the 7-line Previous-version chain (v0.43.0 → v0.42.2 → ... → v0.39.0) with a 1-line pointer to BACKLOG.md Quick Stats + CHANGELOG.md for the full release timeline. US60's strategy (drop the chain) is finally executed here.
• SPEC.md line 3 — Status header rewritten: same treatment as README.md; spec-rev bumped v0.19 → v0.20 and app version lead updated v0.43.0 → v0.44.0. The 10-line Previous-version chain (v0.42.2 → v0.38.0) is gone.
• No code changes, no new public surface, no version bump — content-only revision per US42 protocol. Re-rendered live at https://localhost:8001/info after the post-82b69f6 rebuild (image baa3a54a4d35, 863 MB, 2026-07-07 08:24 UTC).

🔭 Vision (previous Unreleased)

US31 closure lands alongside US60 — both content-only revisions (no version bump per US42 protocol). US31 closes a 50-day-old partial ship: original multi-stage restructure shipped as v0.22.0 on 2026-06-10 (commit 2b9ecce, prod image 1.11 GB → 863 MB, devDeps-free, BuildKit cache mounts); the unfulfilled 350 MB on-disk target is explicitly abandoned per JP's 2026-06-30 inspection — marginal 513 MB savings on a single cached image doesn't justify a 1-day pnpm deploy refactor or an architectural rewrite (lazy-load three+globe.gl for /explore). US60 (shipped earlier today) cleared the Status-paragraph drift class in the same session.

📝 Documentation (US31 closure)

• BACKLOG.md US31 Status field — flipped 🚧 Partially shipped → ✅ Closed 2026-06-30 (content-only revision). Documents the original v0.22.0 partial ship (commit 2b9ecce, 2026-06-10), the explicit abandonment of the 350 MB target with rationale (marginal value vs. architectural cost), and the 2 retained known minor nits: (a) 1.2 MB / 146-file /app/src leak in prod runtime — US31 AC fail but dormant (runtime only reads node build, not exploited); (b) test build target exists in the Dockerfile but not wired into a compose service — one stanza of YAML if CI needs it.
• BACKLOG.md US31 DoD section — rewritten as the closure checklist: original v0.22.0 partial-ship ACs checked (multi-stage split, BuildKit cache, devDeps-free, prod healthy, 4 routes 200), 350 MB target explicitly abandoned with documented decision, source-tree leak acknowledged as known nit, test target acknowledged as un-wired, Quick Stats updated, Recent ships entry added, CHANGELOG entry added.
• BACKLOG.md US31 planning-list entry — updated: "partially shipped" status note replaced with "closed 2026-06-30 as content-only revision" note, documenting the abandonment and the 2 retained nits.
• BACKLOG.md Quick Stats pipeline bullet — "1 in progress (US31 ... 350 MB target remaining)" → "0 in progress (US31 closed 2026-06-30 as content-only revision — original multi-stage ship landed as v0.22.0 on 2026-06-10, 350 MB follow-up target abandoned per JP's inspection)".
• BACKLOG.md Recent ships — new #### US31 — closed 2026-06-30 (content-only revision) H4 entry at the top of Recent ships, summarising the 50-day arc (partial ship at v0.22.0 → 350 MB follow-up abandoned on 2026-06-30 → US closure).
• BACKLOG.md US31 detail block "Closure rationale" subsection — added 2 new bullets after "What this DOES pre-empt": (i) what the US actually closed with (863 MB as final state, the 2 known nits documented, the architectural cost to hit 350 MB); (ii) why this closure pattern matters (US31 sat "1 in progress" for 50 days, closing it explicitly converts an abandoned target into a documented decision for the next agent / future JP).
• BACKLOG.md Last Updated line — bumped to 2026-06-30 15:55 UTC with the US31 closure summary.
• No code changes, no new public surface, no version bump — content-only revision per US42 protocol (same pattern as US60 / US57 / US49).
• Verification: pnpm check 0/0, pnpm lint clean, pnpm prettier --write BACKLOG.md no changes needed (file already formatted). The BACKLOG content will be live after the next prod rebuild (BACKLOG.md is baked into the image and rendered live via /info).

📝 Documentation (US60)

• SPEC.md line 3 — Status line rewritten: dropped the 13 KB giant roll-up previous-version chain that claimed **53 US shipped** (canonical is 55); kept the last 5 versions inline (v0.39.0 → v0.43.0) with their test-count chain; added a 1-line pointer to BACKLOG.md Quick Stats as the canonical source + a 1-line pointer to CHANGELOG.md for older ship history. Spec-rev bumped to v0.20 (2026-06-30).
• README.md line 5 — Status line rewritten: same treatment as SPEC.md line 3; dropped the 15 KB roll-up chain that claimed **50 US shipped** (the worst offender — also built up a "Latest ship: US51" block inside the chain which compounded the staleness); kept the last 5 versions inline + pointer to CHANGELOG.md / BACKLOG.md Quick Stats.
• BACKLOG.md line 728 — US22 planning-list checkbox flipped [ ] → [x] with a 1-line status note: "Status 2026-06-20: shipped (doc-only — 6-lens design review; output consumed by US51)".
• BACKLOG.md line 1464 — Second US22 planning-list checkbox (the Sprint 5+ cross-cutting list) flipped [ ] → [x] with the same status note.
• BACKLOG.md line 733 — US31 planning-list inline size claim updated from "1.11 GB → 680 MB (39% smaller)" to "1.11 GB → 863 MB (per 2026-06-26 measurement)" — the 680 MB was an early partial-ship measurement before US34's BuildKit cache mounts + US46's build memory fix landed (both increased on-disk size to 863 MB but with much better rebuild speed + reproducibility); retroactively updated to match Quick Stats line 448.

[0.44.0] - 2026-06-30

🔭 Vision

US58 — Investor-facing deck + /deck page for cold outreach. A new public surface (/deck route, 12 slides, print-friendly via @media print) gives JP a single shareable link that walks an investor through problem → solution → traction → team → ask → use of funds → comparable rounds → honest unknowns + CTA without requiring them to navigate the whole site. All copy is sourced verbatim from the canonical /investors page (US29 + US51) — no new claims invented. The same route renders cleanly on phone + desktop + print; "Download PDF" button triggers window.print() so the investor can save as PDF from their browser (no PDF generator, no new npm deps). 550 → 562 unit tests green, +12 US58 assertions. Minor bump per semver protocol — new public surface.

✨ Added (US58)

• src/routes/deck/+page.svelte — New 12-slide investor deck. Title / Problem / Solution / Why this / Why now / Why us / Market (TAM-SAM-SOM stat cards) / Traction (4 stat cards) / Team / Ask + Use of Funds (5-line breakdown) / Comparable rounds (Starcloud + US competitors) / Honest unknowns + CTA (mailto:). All copy sourced verbatim from /investors (US29 + US51) — no new claims invented. Tailwind classes only, no new npm deps.
• src/routes/deck/+page.server.ts — Empty server load (per US58 AC: route is purely static).
• src/routes/deck/deck.test.ts — 12 TDD-shaped assertions: page exists, server load exists, data-testid="deck-page" wrapper, "Deck" in <title>, exactly 12 data-testid="deck-slide-*" sections, "Download PDF" button, window.print() wired up, mailto CTA, TAM/SAM/SOM stat cards present, @media print + page-break-after rules, deck-print-hide hides nav header on print, canonical /investors content referenced (Three Whys, €2-5M ask, TAM/SAM/SOM numbers, Starcloud comparable).
• src/routes/deck/+page.svelte <style> — @media print rules: 1 slide per A4 page (page-break-after), deck-print-hide hides nav header + Download button on print, :global(body) background flips to white, slide text colors flip to black, stat-card backgrounds flip to white. Slide CSS for the on-screen layout: 100vh-per-slide on screen, stat-grid responsive (auto-fit minmax(220px, 1fr)).
• src/lib/components/nav/Nav.svelte — Added /deck to the nav items array, placed between /competitors and /info (primary audience route before meta route). Active-state highlighting inherits from the existing isActive() helper.

[0.43.0] - 2026-06-29

🔭 Vision

US59 — "Why we win" callout on /competitors. The page had 11 verified competitor profiles and a tag-chart cross-section but no high-machines positioning — an investor landing on /competitors got a data dump but not the answer to the silent "why should I care?" question. US59 adds a 4-card bento section between the page header and the existing "Competitors by tag" chart, explaining our actual competitive lane vs. the 11 profiled companies. The 4 cards are sourced verbatim from the /investors "Why us" prose (US29 + US51): AI-native engineering (deep RL for mission-architecture design — AI4CE PhD, 6 papers, 1 ESA CDF collaboration; a capability 30-yr primes can't replicate), Venture speed (primes run 5–10 yr cycles vs. our mid-2027 architecture window), Primes as channel (their slow cycles turn them into integrators/acquirers once the standard is locked — we race for the standard, not the compute market), Open methodology (open research loops outpace closed prime culture ~5×; same reason Linux won the OS war). Visual pattern mirrors US54 home page Three Whys (BentoGrid + 4 BentoCell size="md", 2×2 on lg+, 1 per row on mobile) + title + headline (sky-300 accent) + 2-3-line body. Framing subtitle above the cards: "These 11 companies are not our real competitors. Here's our actual lane." Footer links to /story (Three Whys long-form) and /investors (full pitch + ask). All copy is verbatim from /investors — no new claims invented. Plan-check pass: the original US59 ACs contained hallucinated tech claims (Phi-3 Mini reasoning in 1MB / Terraform-for-space infra-as-code) from the fallback model — those were replaced with the real /investors wedge per JP's "I follow your suggestions" reply (2026-06-29 23:12 UTC); commit caed1d8 rewrote the AC block in BACKLOG, commit e6f2985 shipped the corrected implementation. 550 unit tests green (+4 US59 assertions: section-exists + BentoGrid + each-loop + 4-card-s + footer-link targets); pnpm check 0/0; pnpm lint clean. Live at https://localhost:8001/competitors verified via the curl + JSON-unescape pattern from MEMORY lesson #19. Minor bump per semver protocol — new user-facing surface.

✨ Added (US59)

• "Why we win" callout section on /competitors (src/routes/competitors/+page.svelte) — 4 BentoCells in a BentoGrid above the existing "Competitors by tag" chart, each with title + headline (sky-300 accent) + body (~2 lines):
  • AI-native engineering — Deep RL for mission-architecture design. Our AI4CE research (6 papers, 1 ESA CDF collaboration) applies deep reinforcement learning to the architecture-design loop — a capability 30-year engineering cultures cannot replicate.
  • Venture speed — 18 months is the window. Primes run 5–10 year mission cycles. The architecture window for orbital compute closes mid-2027. Whoever locks the standard then sets the bar the incumbents have to clear.
  • Primes as channel — Slow primes = our distribution. Once the standard is set, their slow cycles turn the primes into integrators and acquirers — not competitors. We race for the standard, not the compute market.
  • Open methodology — Open beats closed on iteration speed. Open research loops outpace closed prime culture ~5× on iteration speed. The agent framework is open-stack; theirs is closed. Same reason Linux won the OS war.
• Framing subtitle above the 4 cards: "These 11 companies are not our real competitors. Here's our actual lane." — flips the page from descriptive (what we tracked) to positioning (why we win).
• Footer paragraph linking to /story (Three Whys long-form) and /investors (full pitch + ask) — readers who want the deep context can find it without leaving the family of pages.
• Stable data-slug per card (ai-native / venture-speed / primes-as-channel / open-methodology) — supports per-card scroll anchors (/competitors#why-we-win-ai-native) and regression tests.

🔧 Changed (US59 ship)

• SPEC.md (line 3, Status header) — bumped v0.42.2 → v0.43.0; new parenthetical describes US59; old v0.42.2 (US56) appended to the "Previous version" chain. Also fixed two pre-existing stale v0.43.0 mentions for US54 (line 79 + line 106) which claimed US54 shipped as v0.43.0 — actually US54 ships as v0.42.0 per CHANGELOG/BACKLOG/package.json; v0.43.0 is now correctly claimed by US59. Per MEMORY lesson #20 (sweep the read-only parts now): same-commit fix prevents future drift that would have been caught on the next audit.
• README.md (line 5, Status header) — same Status header bump as SPEC.md; the v0.42.2 (US56) chain now goes v0.43.0 → v0.42.2 → v0.42.1 → v0.42.0. Ship count: 54 → 55 US shipped; 546 → 550 unit tests green.
• BACKLOG.md — US59 entries ticked ([ ] → [x] for all 7 ACs + all 11 tasks + all 6 DoD bullets); Last Updated line bumped to 2026-06-29 23:35 UTC; "Last US added to planning" now points to US58 (US59 promoted to shipped); "Last US shipped" line bumped from US56 (v0.42.2, 2026-06-27) → US59 (v0.43.0, 2026-06-29); Quick Stats line + bullet list updated (54 → 55 shipped, 2 in planning → 1 in planning (US58), 546 → 550 tests, pipeline 59 → 61 total); new "Recent ships" #### US59 — v0.43.0 (2026-06-29) H4 entry added at the top.
• package.json — version 0.42.2 → 0.43.0.

📦 Technical details (US59)

• 1 page-template edit: src/routes/competitors/+page.svelte — added whyWeWin typed ReadonlyArray of 4 cards above <svelte:head> (with slug typed string-literal union ai-native | venture-speed | primes-as-channel | open-methodology); added the new <section data-testid="competitors-why-we-win"> block inside {:else} between the page header and the existing tag-chart <BentoCell size="full">; BentoGrid + 4 <BentoCell size="md" id="why-we-win-${w.slug}"> (2×2 on lg+ via lg:col-span-2, single-column on mobile via the US7 responsive pattern); footer <p> with 2 inline <a> links (href="/story" and href="/investors").
• 1 test extension: src/routes/competitors/competitors.test.ts — new describe("US59 — /competitors 'Why we win' callout") block with 4 assertions: (a) data-testid="competitors-why-we-win" section exists, (b) the section wraps a <BentoGrid> with {#each whyWeWin as w (w.slug)} keyed loop, (c) all 4 expected slugs are declared in the whyWeWin typed array literal (each tested individually + the each-loop is keyed by slug so each card renders the right DOM id), (d) the callout footer has both why-we-win-link-story + why-we-win-link-investors links pointing to /story and /investors respectively.
• Verbatim copy from /investors — all 4 card bodies are sourced verbatim from the /investors "Why us" prose (US29 + US51). Any future copy change on /investors propagates to /competitors via copy update + a single commit.
• Lesson-#20 sweep — the US54 v0.43.0 mistake in SPEC.md would have surfaced on the next verification-pass audit as drift between code reality and doc text; fixed in this same commit before it became that drift. The "50 US shipped" / "53 US shipped" / "54 US shipped" count in the giant Status paragraph (line 3) was not updated to "55" — that's a doc refactor outside US59's read-only scope and got the same drift-deferred-to-follow-up treatment as MEMORY lesson #0a / lesson #20 recommended (sweep small things, open a US for storytelling-debt rewrites). Tracked separately.
• Tests + checks: 550/550 unit tests green (+4 US59 assertions from 546); pnpm check 0/0 errors and 0/0 warnings; pnpm lint clean (one Prettier formatting auto-format folded into the prior docs commit caed1d8).

⚠️ Known Issues (US59 ship)

• The BACKLOG.md giant Status paragraph ship-count claims are still stale in places — the line-3 Status chain + Quick Stats bullets got the canonical "55 US shipped" + "550 tests" updates, but the super-long compounded Status chain at the very top of the file still claims "53 US shipped" / "54 US shipped" in places. Sweep them in a follow-up pass per MEMORY lesson #20 (open a US for the doc-counter sweep; not in scope for US59).
• /deck page (US58) still pending — the only remaining "in planning" US after this ship. Co-planned with US59 on 2026-06-29 19:10 UTC; the same plan-check pass corrected the US58 AC #2 "traction" line to reference US1–US57 instead of the stale US1–US56. US58 ships separately per JP's sequencing.

[0.42.2] - 2026-06-27 On 2026-06-29, the stale-Status audit (commits 41b2cdd + 828b9f5 + a22e509) fixed 7 of 7+ lesson-#0a siblings (6 Status fields + line 439 US55 mention + Last Updated line + Quick Stats block). The "Next sprint (5+)" line on BACKLOG.md line 439 was the one remaining design-decision part: it still named US22 (design review — shipped 2026-06-21 as v0.39.0 via US51), US28 (competitor profile detail pages — shipped v0.25.0), and US29 (Three Whys narrative deconstruction — shipped v0.34.0) as "Next sprint", but all three actually shipped weeks/months ago. Same lesson-#0a root cause as the Status field drift; the difference is that the line is a curated forward-looking list (not a count) so the rewrite is a design decision per lesson #20 — not a sweep, a US. The line is now: "⏭️ Next sprint (5+): US19 (a11y audit), US20 (SEO + meta tags), US21 (quarterly content refresh), US48 (real DNS + Let's Encrypt, blocked on JP's DNS step) — cross-cutting housekeeping. US30 (hub-and-spoke design fully closed 2026-06-19) — Phases 1–4 shipped as v0.35.0 / v0.36.0 / v0.37.0 / v0.38.0, all 4 audience cards now routable. US31 (Docker build runtime inspection & optimisation) — in progress (multi-stage Dockerfile shipped, 350 MB target remaining with pnpm deploy)." — the 4 actual future-sprint USes (US19/20/21/48) are now front-and-centre; the 2 design-history USes (US30 closed, US31 in-progress) keep their context. Quick Stats line 428 "1 in planning" → "0 in planning". Same-cycle CHANGELOG hygiene: the US56 content that was orphaned in [Unreleased] (because US56 shipped as v0.42.2 on 2026-06-27 but never got moved to a dated block) is now in a new [0.42.2] - 2026-06-27 section, leaving [Unreleased] for the next ship. No version bump per US42 protocol — content-only refresh, no new user-facing surface. Production container live at https://localhost:8001/info showing the rewritten line + the cleaned [Unreleased]/dated split.

✨ Added (US57)

• "Next sprint (5+)" line in BACKLOG.md rewritten (line 439 of the Quick Stats block) — the 3 stale references to shipped USes (US22 / US28 / US29) are gone; the 4 actual future-sprint USes (US19 a11y / US20 SEO / US21 quarterly / US48 DNS) are the headline. US30 (hub-and-spoke design fully closed 2026-06-19) + US31 (Docker build runtime inspection & optimisation, in progress) keep their design-history context. Same pattern as the US38 audit-restructure: the line now reads as a curated forward-looking list, not a stale "what's coming" wall.
• CHANGELOG hygiene (US57 side-effect) (CHANGELOG.md) — the [Unreleased] block still contained US56's full content (Vision / ✨ Added / 🔧 Changed) because US56 shipped as v0.42.2 on 2026-06-27 but the [Unreleased] → [0.42.2] move was never made. Moved: US56 content is now in a new [0.42.2] - 2026-06-27 block, and [Unreleased] is clean for the next ship. Same lesson-#0a root cause: the date wasn't updated when the version was finalised. The corrected structure mirrors the US56 closeout's own note: "The corrected [Unreleased] section is now split cleanly into a [0.42.1] - 2026-06-27 block (with the US55 details, US49 content-only refresh, and the duplicate US55 Changed section removed) and a new [Unreleased] block (with the US56 details for the v0.42.2 ship)" — US57 closes the loop by promoting the US56 [Unreleased] content into its proper dated section.
• Quick Stats "1 in planning" → "0 in planning" (BACKLOG.md line 428) — US57 was the only "in planning" US; on ship, the count drops to 0. The bullet structure (54 shipped + 0 in planning + 4 future sprint) is back to the pre-US57 form, matching the "Last US added to planning" header (which still references US57 as the most recent planning entry added, regardless of ship status).

⚠️ Known Issues (US57)

• US57 has no version bump and no [Unreleased] from [0.42.2] split commit hash — per US42 protocol, content-only refreshes (no new feature, no new public surface) get no version bump; they go in [Unreleased] in CHANGELOG.md, not in a new dated version. The ship commit (this one) is the canonical reference; the git log + commit message is the only way to find the US57 content. If JP wants a v0.42.3 patch bump for traceability, the change is a one-liner in package.json + a US58 closeout commit. Out of scope for this US.
• The BACKLOG.md "Last Updated" line + "Last US added to planning" line drift if not maintained — same lesson-#18 same-commit tax. US57's ship commit (this one) includes the Last Updated bump in the same diff. Future USes should follow the same pattern; the existing lesson #18 + lesson #16 in MEMORY.md document the rule.

[0.42.2] - 2026-06-27

🔭 Vision

US56 — Render competitor profile markdown body as structured system overviews. On 2026-06-27, JP was looking at the live /competitors/starcloud page and said "I want full overviews about the system in dedicated sections" — the rich hardware deep-dives (Starcloud-1's 60 kg H100 demonstrator with Karpathy's nano-GPT, Starcloud-2's 8 kW Blackwell-B200 commercial platform with Crusoe Cloud + AWS Outposts) were sitting in the research markdown, not on the website. The current page only shows the curated digest (summary / highlights / watchlist / sources); visitors who wanted system-level specifics had to click through to GitLab. Two paths to fix: (a) lift the hardware sections from the markdown to the page as dedicated cards (lowest-complexity win, reuses the existing bind-mount, no new routes), (b) build a new Systems tab + nested routes (over-engineered for content that's likely to stay flat). Path (a) wins. Implementation: at request time, the server load reads the markdown body, splits it on H2 boundaries (respects fenced code blocks so the Starcloud-1/2 hardware tables don't get sliced), filters H2s via a heuristic (separator — or -, OR a known system-keyword prefix like starcloud / bluebeam / cubesat), renders each matched section via the existing marked instance (the US41 mermaid passthrough is shared via a new $lib/utils/markdownRenderer util so /tech and the new /competitors/[slug] use the same renderer), and passes the array of { heading, html } to the page. The page renders an optional <section data-testid="system-overviews"> between the digest and the "Source on GitLab" footer — when systemSections.length === 0 the section wrapper is omitted entirely (graceful degradation, profiles like axiom-space with no system H2s render no extra block). The markdown remains the single source of truth per US55 + the US6 bind-mount: editing a system section heading or body on the host shows up on the live page on the next reload, with no rebuild. Two new shared utils ($lib/utils/markdownSections + $lib/utils/markdownRenderer); one new readMarkdownBody helper in $lib/competitors/data (mirrors the US55 readMarkdownUpdated shape); the marked.use({ renderer }) setup is lifted from /tech so the global marked instance is configured exactly once. 546 unit tests green (+19 US56 assertions: 16 in markdownSections.test.ts + 3 in [slug]/[slug].test.ts). Minor bump per semver protocol — new user-facing surface (system-overview cards on the detail page). Ships as v0.42.2.

✨ Added (US56)

• System overview cards on /competitors/[slug] (src/routes/competitors/[slug]/+page.svelte, src/routes/competitors/[slug]/+page.server.ts, src/lib/utils/markdownSections.ts, src/lib/utils/markdownRenderer.ts, src/lib/competitors/data.ts) — the research markdown's "system overview" H2 sections (e.g. ## Starcloud-1 — Technology Demonstrator and ## Starcloud-2 — Commercial Platform in 01-research/competitor-profiles/starcloud.md) are now rendered as dedicated content cards under the typed digest. Each card shows the section heading + the rendered markdown body (tables, lists, prose). The page is now progressive-disclosure: the typed digest is the 1-paragraph view, the system cards are the deep dive, the "Source on GitLab" link is the full research notebook. Profiles without system sections (e.g. axiom-space, eric-schmidt-relativity) render no extra block — systemSections.length === 0 skips the section wrapper entirely (graceful degradation). The markdown remains the single source of truth per US55 + the US6 bind-mount: editing a system section heading or body on the host shows up on the live page on the next reload, with no rebuild.
• New $lib/utils/markdownSections util — parseSections(body) splits a markdown body on H2 boundaries (respects fenced code blocks, strips leading + trailing blank lines from each section); isSystemSection(heading) returns true for headings that look like a named hardware platform (heuristic: contains an em-dash or hyphen with spaces, OR starts with a known system-keyword — starcloud / bluebeam / axiom dc / axdcu / cubesat / cdpu); parseSystemSections(body) is the convenience combo. Locked in by 16 unit tests in markdownSections.test.ts (split on H2, ignore H1/H3-H6, preserve fenced code blocks spanning H2s, handle both ```and ~~~ fences, real-file regression test onstarcloud.md+ graceful-degradation test onaxiom-space.md).
• New $lib/utils/markdownRenderer util — the marked.use({ renderer }) setup with the US41 mermaid passthrough was lifted from src/routes/tech/+page.server.ts into a shared module so /competitors/[slug] can reuse the same mermaid-aware renderer without re-installing it (two marked.use calls on the global marked singleton would clobber each other). The shared util exports renderMarkdown(markdown) (async, returns the HTML string) and re-exports the configured marked instance. Future routes that render markdown body content can import from here too.
• readMarkdownBody(slug) in $lib/competitors/data.ts — mirrors the US55 readMarkdownUpdated(slug) shape: same COMPETITOR_PROFILES_DIR env override, same async node:fs read, same parseFrontmatter import, returns the body (everything after the closing --- of the frontmatter block) or undefined if the file is missing / unreadable / has no frontmatter. The detail-page server load calls it once per request, then runs the body through parseSystemSections and renders each section via the shared renderMarkdown util.
• Progressive disclosure: the /competitors/[slug] page stack is now 4 layers — the header (name, oneLiner, tags, freshness badge per US55), the typed digest (summary / highlights / watchlist / sources — unchanged), the new system-overview cards (US56), and the "Source on GitLab" footer link. Each layer is independent: a profile with no system sections still has a complete page; a future profile with 5 system sections renders 5 cards without changing the digest.

🔧 Changed (US56)

• /tech route imports the shared markdown renderer (src/routes/tech/+page.server.ts) — the local marked.use({ renderer }) setup + the US41 mermaid passthrough are gone from this file; the route now imports renderMarkdown from $lib/utils/markdownRenderer. Behaviour is unchanged: every /tech section that contains a ```mermaid block still renders as <pre class="mermaid">…</pre>(the form that mermaid 11.x picks up viaquerySelector). The change is internal — no visible difference to visitors.
• mermaid.test.ts and tech.test.ts updated to lock in the new shared location: both tests now check that $lib/utils/markdownRenderer.ts contains the <pre class="mermaid"> string and the renderer.code override, and that /tech imports from the shared util (and does NOT call marked.use locally — that would clobber the global instance).
• 🪛 Fixed (US56) — duplicate ### 🔧 Changed (US55) block in [Unreleased] — caught while writing the US56 entry. The duplicate was a copy-paste artefact from the US55 closeout on 2026-06-27 (the section was appended twice to [Unreleased]). Removed in this closeout. The corrected [Unreleased] section is now split cleanly into a [0.42.1] - 2026-06-27 block (with the US55 details, US49 content-only refresh, and the duplicate US55 Changed section removed) and a new [Unreleased] block (with the US56 details for the v0.42.2 ship).

• 2026-06-27

🔭 Vision

US55 — Single source of truth for competitor profile dates (markdown frontmatter ↔ data.ts). On 2026-06-27, JP noticed the "Reviewed YYYY-MM-DD" badge on /competitors/starcloud (and the other 10 profiles) still showed 2026-06-20 even though US49 had refreshed the markdown updated: frontmatter to 2026-06-26 the day before. Root cause: the date was duplicated in two places — the markdown frontmatter AND the bundled CompetitorProfile shape in src/lib/competitors/data.ts. US49 only refreshed the markdown, leaving data.ts stale; the badge only updated after commit 825d677 + a full prod container rebuild (the OOM-prone path per memory §14, ~5 min + watch the gateway watchdog). Two sources of truth = guaranteed drift on the next refresh. Fix: read the date from the markdown at request time via the existing US6 bind-mount of 01-research/competitor-profiles/ → COMPETITOR_PROFILES_DIR in the container. Remove the duplicated updated: field from the CompetitorProfile interface + the 11 entries in data.ts. Extract the hand-rolled parseFrontmatter from src/routes/tech/+page.server.ts (US32 + US41) into a shared $lib/utils/frontmatter util so both routes can read the frontmatter without re-implementing it (no new deps, no new build step). Patch bump per semver protocol — bug-class fix, no new user-facing surface (the badge text + position are unchanged). Ships alongside US49 as v0.42.1.

US49 — Competitor profile verification & refresh (April–June 2026). All 11 competitor profiles in 01-research/competitor-profiles/ updated with fresh milestones from the last 3 months. Content-only — no version bump per US49 acceptance criteria ("this is content-only; the site reads the profiles via bind-mount so changes are live immediately on next deploy").

✨ Added

• Competitor profile updates (01-space-AI/01-research/competitor-profiles/*.md) — all 11 profiles refreshed with April–June 2026 milestones:
  • Batch A (2026-06-20): mantis-space (New Mexico incentives, stealth exit), agnikul-cosmos (semi-cryogenic engine test, NeevCloud partnership), neevcloud (Agnikul MoU, India's first indigenous ODC)
  • Batch B (2026-06-26): axiom-space ($525M oversubscribed round + Texas redomicile + Swiss sub + Prada LCVG + NASA core-module abandonment), aetherflux (Cowboy Space rebrand + $275M Series B + 20,000-sat Stampede FCC filing + end-2028 rocket target), alibaba-qwen-departures (two-track narrative: Alibaba closes flagship weights + Lin Junyang $2B pre-seed AI lab)
  • Batch C (2026-06-26): starcloud (Oct 2026 → Jan 2027 launch slip + Series A co-leader fix Benchmark + EQT + $200M total funding + Starlink Mini Laser deal + Series B talks at $2.2B), blue-origin (May 28 New Glenn explosion + May 5 NASA objection + Isaacman 2028 timeline + threat downgrade HIGH→MEDIUM), eric-schmidt-relativity (Aeolus reframed as NASA science contract + Terran R Stennis milestone + instrument suite)
  • Batch D (2026-06-26): spacex (IPO completed June 12: $75B raised, >$2T valuation, Musk first trillionaire + Anthropic date correction May 6 + Starship Flight 12 partial success + FAA mishap May 27 + Blue Origin explosion May 29 + Starlink 12,342 launched + AI1 + Gigasat + Google $30B compute deal), google-suncatcher (Trillium TPU radiation test corrected to Nov 2025 + clarified $30B Google-SpaceX deal is Nvidia cloud compute, not Suncatcher)
• 41 new sources added to sources.db (IDs 366–410) — SpaceNews, Reuters, Bloomberg, NYT, TechCrunch, SpaceX primary, Wikipedia, Forbes, NPR, CNBC, Spaceflight Now, KeepTrack.space, SatNews, Via Satellite, BusinessWire, Axiom Space primary, University of Houston, Texas Governor, xAI, Decrypt, 36Kr, GuruFocus, BigGo Finance, The Information, Caixin Global, QwenLM GitHub.
• No website code changes — content-only refresh, live immediately via bind-mount.

🔧 Changed (US55)

• Single source of truth for competitor profile dates (src/lib/competitors/data.ts, src/routes/competitors/[slug]/+page.server.ts, src/routes/competitors/+page.server.ts) — the CompetitorProfile interface no longer carries an updated: field; the 11 hand-written entries no longer duplicate the date. The "Reviewed YYYY-MM-DD" badge on /competitors/[slug] and the <time datetime="…"> element on /competitors now both read the date from the markdown's updated: frontmatter at request time (via the US6 bind-mount). Editing the markdown's updated: field on the host is reflected on the live page on the next reload, without a rebuild.
• Shared parseFrontmatter util (src/lib/utils/frontmatter.ts) — the hand-rolled YAML-ish parser (originally shipped in US32, extended in US41 for folded / literal block scalars) is now extracted from src/routes/tech/+page.server.ts into a shared module. Both /tech (the original consumer) and $lib/competitors/data.ts (the new consumer for updated: reads) import from $lib/utils/frontmatter. No new npm deps. The shared parser is locked in by src/lib/utils/frontmatter.test.ts (9 new tests covering scalar pairs, bracketed lists, quoted strings, folded / literal block scalars, missing / malformed frontmatter, and the real starcloud.md shape that drove US55).
• Frontmatter added to alibaba-qwen-departures.md — previously the only profile without a ----delimited frontmatter block (it used inline **Date:** + last updated in the body). Added the standard title: / date: / updated: / tags: / status: frontmatter so the new "read from markdown" path picks it up. tags now includes competitor (was untagged before; the bar chart count moves from 10 → 11 — see competitor-tags.test.ts).
• CompetitorProfile shape slimmed from 12 fields to 11 — updated: string removed. Call sites updated: [slug]/+page.server.ts calls await readMarkdownUpdated(slug) and passes the result to formatReviewedDate(); the list-page server load calls Promise.all(profiles.map(readMarkdownUpdated)) and attaches the date to each CompetitorListEntry for the <time datetime="…"> semantic-HTML element.
• New regression test in [slug]/[slug].test.ts — locks in the US55 contract: the bundled shape carries no updated: field on the 11 entries, and the server load derives the date from the markdown via readMarkdownUpdated.
• Defense against re-introduction of the drift bug — the data.test.ts "every entry has the 12 required fields" test is now "every entry has the 11 required fields" (US55 made it 11, was 12 before with updated). The [slug]/[slug].test.ts runtime guard regex catches anyone re-adding updated: "YYYY-MM-DD", on any entry object. The [slug]/[slug].test.ts "the page must NOT read the markdown file system" assertion is updated to "the page must NOT import node:fs directly" — the read is now delegated to readMarkdownUpdated in $lib/competitors/data.

[0.42.1] - 2026-06-27

🔭 Vision

US55 — Single source of truth for competitor profile dates (markdown frontmatter ↔ data.ts). On 2026-06-27, JP noticed the "Reviewed YYYY-MM-DD" badge on /competitors/starcloud (and the other 10 profiles) still showed 2026-06-20 even though US49 had refreshed the markdown updated: frontmatter to 2026-06-26 the day before. Root cause: the date was duplicated in two places — the markdown frontmatter AND the bundled CompetitorProfile shape in src/lib/competitors/data.ts. US49 only refreshed the markdown, leaving data.ts stale; the badge only updated after commit 825d677 + a full prod container rebuild (the OOM-prone path per memory §14, ~5 min + watch the gateway watchdog). Two sources of truth = guaranteed drift on the next refresh. Fix: read the date from the markdown at request time via the existing US6 bind-mount of 01-research/competitor-profiles/ → COMPETITOR_PROFILES_DIR in the container. Remove the duplicated updated: field from the CompetitorProfile interface + the 11 entries in data.ts. Extract the hand-rolled parseFrontmatter from src/routes/tech/+page.server.ts (US32 + US41) into a shared $lib/utils/frontmatter util so both routes can read the frontmatter without re-implementing it (no new deps, no new build step). Patch bump per semver protocol — bug-class fix, no new user-facing surface (the badge text + position are unchanged). Ships alongside US49 as v0.42.1.

US49 — Competitor profile verification & refresh (April–June 2026). All 11 competitor profiles in 01-research/competitor-profiles/ updated with fresh milestones from the last 3 months. Content-only — no version bump per US49 acceptance criteria ("this is content-only; the site reads the profiles via bind-mount so changes are live immediately on next deploy").

[0.42.0] - 2026-06-25

🔭 Vision

US54 — Pitch-led /home landing redesign. JP said "I am not yet happy with the landing page" on 2026-06-25 during the house-cleaning session (after the 4 GB swap + prod redeploy + Docker/caches cleanup). The diagnosis: the existing +page.svelte had an 80vh 3D-earth hero with a generic subtitle ("the European orbital data centre idea") and two content-discovery CTAs (/explore + /story); the pitch (Three Whys, target architecture, status / ask) was buried on the audience pages, never visible to a casual visitor who didn't click further. The deep-tech seed-startup standard (Starcloud, Axiom, Anduril, Commonwealth Fusion all lead with the pitch, not with the product animation) is: 5-section pitch-led flow → audience routing. US54 replaces the landing with a 6-section pitch-led flow: 40vh hero (3D earth demoted to background texture) with ONE primary CTA "Read the pitch →" pointing to /investors; existing OrbitComparison (US10) sits between hero and pitch; Three Whys preview (3 bento cells with compressed copy from /story US29 + US51); The Number strip (120 sats · 600 km · 100 TFLOPS/sat · 5 EU ground stations, matches DEFAULT_PARAMS in src/lib/explore/params.ts and the /investors "Our target architecture" prose — consistency check is T54.4 acceptance); Status / ask strip (4 chips: Seed stage · Raising €2–5M · Architecture lock phase · ~24 months to Series A); the existing "Who are you?" hub (US30 Phase 1) stays as the closing block. Minor bump per semver protocol (new landing layout + new content surface; 517 unit tests green, +0 US54 assertions — pure layout + copy change, no new code paths; Button.test.ts updated to check for the new single /investors CTA instead of the old Explore + Story pair).

✨ Added

• Three Whys preview section (src/routes/+page.svelte) — 3 bento cells with the 30-second version of the Three Whys from /story (US29 + US51): Why this (orbital compute thesis — continuous solar + radiative cooling + no permitting), Why now (18-month architecture window — LEO economics flipped + AI inference 10× scaled + standards-setting race), Why us (6 AI4CE papers + PhD in mission-architecture design + the only open-stack agent framework for concurrent engineering in space). Each cell has a sky-coloured headline + 2–3 lines of compressed body copy. Footer link "Read the full Three Whys on /story →" routes the visitor to the long-form page.
• The Number strip (src/routes/+page.svelte) — single bold line with 4 numbers: 120 satellites · 600 km LEO · 100 TFLOPS/sat · 5 EU ground stations. The numbers are duplicated from DEFAULT_PARAMS in src/lib/explore/params.ts (US51 target architecture) for the /, /investors, /explore consistency. The 4 numbers mirror Starcloud's "gigawatts" framing — concrete scale, no paragraphs of text.
• Status / ask strip (src/routes/+page.svelte) — 4-chip row with the project's status + ask: 🌱 Seed stage · 💶 Raising €2–5M · 🔒 Architecture lock phase · ⏱ ~24 months to Series A. Mirrors how Anduril + Starcloud + Axiom all surface the funding status on their landing pages. Each chip is an emoji + label pair, no extra components, no new deps.

Changed

• src/routes/+page.svelte — rewritten as a 6-section pitch-led flow. Hero shrunk from h-[80vh] to h-[40vh] (the 3D earth stays, demoted to background texture); overlay trimmed to title + slogan + one-liner + ONE primary CTA ("Read the pitch →" linking to /investors); the old dual CTAs (/explore + /story) are gone from the hero (still accessible via the nav menu). New sections added in order: OrbitComparison (kept from US10) → Three Whys preview → The Number strip → Status / ask strip → "Who are you?" hub (kept from US30 Phase 1). The hub's audiences array still has live: true on all 4 entries + the {#if}/{:else} guard so the existing landing.test.ts keeps passing (US30's "Coming next" branch stays in source, unreachable — keeps the regression gate for the US30 fully-closed state).
• src/lib/components/ui/button/Button.test.ts — updated the static-source checks to match US54's new single CTA: removed the "Explore CTA" + "Story CTA" assertions, added "Read the pitch → CTA pointing to /investors" assertion + "no Explore + Story CTAs in the hero" assertion. Same component, same test file, same 1 test count delta.
• README.md status line — prepended the v0.42.0 ship summary to the existing v0.41.0 / v0.40.0 / v0.39.0 chain.
• SPEC.md — added a "Note on US54 — Pitch-led /home redesign (✅ shipped 2026-06-25, v0.42.0)" block in §1 between the US53 note and the US29 note. Describes the diagnosis (80vh earth + dual CTAs = pitch buried), the fix (6-section flow + single CTA), the CTA decision (/investors over mailto:), the consistency check (numbers match DEFAULT_PARAMS + /investors prose), and the test delta (Button.test.ts updates, no new tests).
• BACKLOG.md — US54 detail block committed in the planning phase earlier today (4878879); the header sections (Last Updated, Last US added to planning, Quick Stats, Current Focus / Next sprint list) were updated when US54 was added.

✅ Verification

• pnpm test:unit — 517/517 green (was 517 at v0.41.0, +0 US54 — pure layout + copy change). landing.test.ts (10 tests) + Button.test.ts (5 tests, 2 updated for US54) green; full project test suite (51 test files / 517 tests) green.
• pnpm check — 0/0 (svelte-check, type-check clean).
• pnpm lint — clean (prettier + eslint pass on the updated +page.svelte + Button.test.ts).
• pnpm build — succeeds on the 3.7 GB agent host now that the 4 GB swap was added 2026-06-25 (the swap unblocks the SSR-finalize OOM that historically hit at the 2.0–2.1 GB peak; with the swap + US46's output.bundleStrategy: 'split', the build runs cleanly in ~2 min).
• Production container rebuild + live verification — deferred to JP's machine per the US46 OOM constraint, now possible on the 3.7 GB agent host with the swap headroom. The +page.svelte change is layout + copy; no new components, no new routes, no new deps; live verification is HTTP 200 on / + visual check that the 6 sections render in order.

⚠️ Known Issues

• CTA target decision — US54's T54.2 prerequisite was a decision: "primary → /investors (deeper pitch) vs mailto: (direct contact)". JP didn't explicitly weigh in; we shipped with primary → /investors because the deep-tech seed-startup standard is pitch-first, contact-second (Starcloud + Axiom + Anduril all do this). If JP wants mailto: instead, the change is one line: href="/investors" → href="mailto:jan-peter@highmachines.space?subject=...". Easy to flip in a follow-up.
• OrbitComparison stays on landing — US54 didn't explicitly address the OrbitComparison component placement. We kept it between the hero and the Three Whys preview (visual tech demo before words). The 6-section flow is dense; if JP wants fewer sections, the cleanest drop is OrbitComparison (it's also accessible via /explore). Tracking as a follow-up decision rather than a US.
• No new tests added — US54 is layout + copy in one file; the pnpm test:unit delta is 0 (517 → 517). The acceptance criterion "Lighthouse performance ≥ 90 on /" (was ~94 before) wasn't measured after the change; the smaller hero should help, but the actual delta is unverified. Run pnpm lighthouse:baseline post-deploy to confirm.

[0.41.0] - 2026-06-22

🔭 Vision

US53 — Customer personas on the 4 audience pages. The audience pages (/customers, /partners, /investors, /regulators) used to describe segments in the abstract — "European hyperscalers", "satellite builders", "deeptech VCs", "EU policy officers" — without showing what success looks like for a specific person in that role. The US22 design review flagged "show me the customer — name the anchor" as a top sceptic push-back. US53 answers with typology, not named prospects: 10 fictional but realistic archetypes (3 customers + 3 partners + 2 investors + 2 regulators), each with name + role + 1-line goal + 1-line pain + 1-line objection + 1-line voice quote, rendered as inline "Who this page is for" bento cells on each audience page. The persona set is reusable across the 30-min call regardless of which prospect shows up, and a hard no-named-prospects guardrail test prevents premature disclosure of the actual EU hyperscaler / sovereign-AI / partner pipeline. The 4-audience hub-and-spoke (US30) now has a face for every visitor. Minor bump per semver protocol (new content surface; 517 unit tests green, +23 US53 assertions — 11 in the personas data module + 12 across the 4 audience page tests).

✨ Added

• src/lib/personas/data.ts — typed data module with the CustomerPersona interface (id + segment + name + role + emoji + accent + goal + pain + objection + voice), the personas array of 10 entries, the getPersonasBySegment(segment) lookup, the getPersona(id) lookup, the containsForbiddenProspectName(input) guardrail helper, and the segmentLabel(segment) display helper. Mirrors the src/lib/competitors/data.ts pattern.
• src/lib/personas/PersonaCard.svelte — reusable 7-field persona card component (12×12 colored avatar circle with the persona's emoji + name + role header + 3-line definition list for goal / pain / objection in muted color-coded labels + voice quote in a left-bordered italic blockquote). ~50 lines. Svelte 5 runes mode.
• src/lib/personas/index.test.ts — 11 unit tests for the data module (10 entries, all required fields, unique kebab-case ids, single-sentence goal/pain/objection, segment counts 3/3/2/2, helper lookups, no-named-prospects guardrail that scans both the in-memory data and the source-file personas array for any of the 9 names from the US22 design review).
• 10 customer personas in 4 segments:
  • 🏢 Maria — CTO at a European hyperscaler (sovereign-inference buyer, 2027 EU AI Act compliance).
  • 🇪🇺 Thomas — AI infrastructure lead at a European sovereign-AI programme (scaling beyond terrestrial GPU supply).
  • 🧪 Sophie — Head of research compute at a European AI lab (burst training, 10K-GPU days/month).
  • 🛰️ Lukas — Satellite bus product manager at a European prime (selling bus + integration to orbital-DC operators).
  • 📡 Aino — Ground station network lead at a Nordic operator (maximising GS utilisation outside polar windows).
  • 🔗 Diego — System integrator at a major aerospace prime (combining systems engineering with EU regulatory liaison).
  • 💰 Eva — Partner at a deeptech VC fund (finding the European orbital-DC winner before the category consolidates).
  • 🏛️ Stefan — Principal at an EU sovereign-AI fund (deploying capital that builds European strategic autonomy).
  • 🇪🇺 Hanna — Policy officer at the European Commission DG-CONNECT (EU AI Act + EU Space Law alignment).
  • 🚀 Mateusz — Programme officer at ESA (ARTES + Horizon Europe Cluster 4 + DLR Orbitall alignment).
• "Who this page is for" bento cells on each of the 4 audience pages — placed right after the hero / before the existing SLA / Three Whys / sovereignty content, so visitors see themselves first. 2-3 personas per page (3 customers / 3 partners / 2 investors / 2 regulators). Same mailto: CTA per US51; personas drive self-identification, not routing.

Changed

• /customers + /partners + /investors + /regulators — each gained a new "Who this page is for" full-width bento cell with the matching personas, placed between the hero header and the existing content bento. The CTAs are unchanged across all 4 pages. The /customers and /partners grids use md:grid-cols-3 (3 personas); the /investors and /regulators grids use md:grid-cols-2 (2 personas). No new route, no new CTA, no audience-segment refactor.

Security / Compliance

• No-named-prospects guardrail test (src/lib/personas/index.test.ts) — asserts that no persona field (or any line in the personas data array) contains any of the 9 names from the US22 design review: Hetzner / OVHcloud / Scaleway / IONOS (EU hyperscalers) / Axiom / Starcloud / Kepler / Suncatcher (US orbital-DC competitors) / OHB (European prime). The test fails the build on any future copy-paste accident that would prematurely disclose JP's actual prospect pipeline.

⚠️ Known Issues

• pnpm build not run in this session — the production build is deferred to JP's machine per the US46 OOM constraint (3.7 GB agent host OOMs at the adapter-node finalize step even with NODE_OPTIONS=--max-old-space-size=1500 + output.bundleStrategy: 'split'). The pnpm test:unit (517/517 green) + pnpm check (0/0) + pnpm lint (clean) pipeline passes on the agent host; the build verification is the same pattern as US51 / US46 / US50.
• Production container not refreshed in this session — the live high-machines-website:prod container is still on v0.40.0. The build + container refresh is deferred to JP's machine (the build OOMs on the 3.7 GB host even with the US46 fixes). Same workflow as US46 / US50 / US51.

[0.40.0] - 2026-06-22

🔭 Vision

US52 — OMM-first orbital-data architecture. The static/tle/ folder name was lying: it contained OMM blocks, not TLE text. The TLE fallback (twoline2satrec) was dead code — a year of CelesTrak data shows no malformed OMMs. US52 deletes both: the folder is renamed to static/orbit-data/ (it now says what the data IS, not what legacy code thought it was), the catalog's inlineOMM field is retired in favour of a dataPath pointing at a JSON file, and the propagator is OMM-only via json2satrec. The architecture now matches reality: OMM is the truth, TLE was a legacy fallback that's no longer needed in the cache. The broken TypeScript codemod (the update-inline-omm script that regular failed on a regex match) is replaced by a trivial fetch + writeFileSync. Manual refresh becomes one command (pnpm omm:update-committed); automated refresh becomes cheap on frequent builds (7-day age check in pnpm omm:fetch skips CelesTrak unless the cache is stale). The 5 weekly "chore: refresh TLE cache" commits stop being a thing. Minor bump per semver protocol (architectural change, 494 unit tests green, +2 US52 assertions — small net delta because US52 retired the parse3LE / mergeOMMAndTLE / propagateTLEToECI / getInlineConstellations test cases and replaced them with new tests for the age check, the --force flag, the catalog dataPath reader, and the runtime OMM-only API).

✨ Added

• static/orbit-data/ directory — 7 JSON files, one per satellite/constellation:
  • 4 committed-OMM files (tracked in git): iss.json, starcloud-1.json, tiangong.json, axiom-1.json. Single-sat fixtures, refreshed manually via pnpm omm:update-committed.
  • 2 prebuild files (gitignored): oneweb.json, kuiper.json. Bulk constellations, refreshed weekly via pnpm omm:fetch with the 7-day age check.
  • 1 runtime-fetched entry: axiom (no static file — fetched on demand from /api/celestrak/stations).
• static/orbit-data/.metadata.json — { "lastFetched": "<ISO8601>", "version": 1 }. Tracks when the prebuild cache was last refreshed. Read by pnpm omm:fetch to drive the age check.
• pnpm omm:fetch (renamed from pnpm tle:fetch, new scripts/fetch-omm.ts) — prebuild fetch for the 2 celestrak-prebuild entries. Reads .metadata.json, skips CelesTrak if lastFetched < 7 days old, otherwise fetches OMM (FORMAT=json only — no TLE) and updates the timestamp. --force flag bypasses the age check for manual overrides. Still wired as the prebuild hook (cheap on frequent builds thanks to the age check).
• pnpm omm:update-committed (renamed from pnpm tle:update-inline, new scripts/update-committed-omm.ts) — plain file I/O. Fetches the latest OMM from CelesTrak for each of the 4 committed-OMM satellites and writes the JSON file directly. The fragile TypeScript codemod is gone — no more regex-on-TS-literal fragility. Catalog-drift detection: the script checks every committed-OMM entry has a NORAD mapping, fails loud on drift.

🔄 Changed

• src/lib/satellites/config/types.ts — the SatelliteConfig type: inlineOMM field removed, dataPath: string field added, 'commited-omm' added to the source union (replaces 'inline-omm').
• All 7 per-satellite config files in src/lib/satellites/config/: inlineOMM stripped, dataPath added, source updated. The catalog entries are now pure metadata (id, name, color, symbol, description, etc.) — the OMM data lives in the JSON files, not the TypeScript.
• src/lib/explore/tle.ts — the propagator is OMM-only. propagatePosition (the TLE/OMM auto-detect wrapper) is replaced by propagateOMM; the twoline2satrec fallback is gone. parse3LE, mergeOMMAndTLE, propagateTLEToECI, TLE_CONSTELLATIONS are all removed (~30 lines deleted).
• src/lib/explore/constellation-data.ts — loadConstellation now reads OMMs from entry.dataPath (the catalog-driven URL). The inline-omm branch (which read entry.inlineOMM) is gone.
• src/routes/api/celestrak/[group]/+server.ts — the runtime CelesTrak proxy is OMM-only. Drops the FORMAT=tle fetch, drops parse3LE + mergeOMMAndTLE. Response shape: { group, source, fetchedAt, count, entries: OMMJsonObject[] }.
• src/lib/satellites/runtime.ts — fetchConstellationAtRuntime returns OMMJsonObject[] (was TLEEntry[]).
• src/lib/satellites/catalog.ts — ConstellationEntry no longer has inlineOMM, gains dataPath. getInlineConstellations() renamed to getCommittedConstellations().
• src/lib/explore/ExploreGlobe.svelte — runtimeTLE: Map<string, TLEEntry[]> renamed to runtimeOMM: Map<string, OMMJsonObject[]>; propagatePosition(tle, ...) calls become propagateOMM(omm, ...). Error message "TLE load failed" → "Orbit-data load failed". Overlay comment "catalog + TLE data" → "catalog + OMM data".
• .gitignore — static/orbit-data/*.json is gitignored, with explicit whitelist for the 4 committed-OMM files + .metadata.json (so they're tracked).
• package.json scripts: tle:fetch → omm:fetch; tle:update-inline → omm:update-committed; prebuild hook calls pnpm omm:fetch.

📚 Docs

• README.md — status line updated with the v0.40.0 entry; tree on line ~144 shows the new static/orbit-data/ directory + the renamed scripts; typo on line 5 fixed ("TLE/OMN-based" → "OMM-based"). The "TLE/OMM-based satellite tracking" reference in the v3 Sprint 3 historical note is left intact (describes the v0.15 architecture accurately).
• SPEC.md — §3 (Build) and §7 (Operations) updated for the OMM-first architecture: prebuild → omm:fetch (was tle:fetch), the build-hook failure logging note points to the new omm:fetch script, and the "Skip prebuild hooks in dev" item now describes the 7-day age check as the cheap-on-frequent-builds mitigation.
• BACKLOG.md — US52 added (this entry). US14, US25, US26, US36, US22 references updated from static/tle/ to static/orbit-data/ (the path-only references; the historical narratives are left as-is to preserve the ship history).

⚠️ Known Issues

• Deviations from the US52 plan (deliberate, documented for the commit): the plan said "7 files" (iss, starcloud-1, tiangong, oneweb, kuiper, axiom, axiom-1) but grouped axiom-1 with axiom as "celestrak-runtime". Axiom-1 is a synthetic test fixture (per US36 T36.17) — there is no real satellite to fetch from CelesTrak, and the e2e suite relies on the committed JSON for deterministic marker placement. So axiom-1 was shipped as commited-omm (with its own JSON), not celestrak-runtime. Net result: 4 committed-OMM files (iss, starcloud-1, tiangong, axiom-1) + 2 prebuild (oneweb, kuiper) + 1 runtime (axiom, no static file). The 7-file count in the plan included axiom-1 as a runtime description; the actual on-disk file count for the 7 catalog entries is 6 JSONs + 1 metadata. If JP wants axiom-1 to be celestrak-runtime instead, the change is one line in src/lib/satellites/config/axiom-1.ts (remove dataPath, change source, add a group) and git rm static/orbit-data/axiom-1.json.
• Container still down (per the 2026-06-21 audit's D-1 finding) — docker compose --profile prod up -d --build was not run as part of US52. The agent host OOMs the build per US50/MEMORY §12, so the prod rebuild is JP's task on his machine. Run after git push + git pull.
• The "TLE fetch" wording in non-project docs (the workspace memory notes, the verification-pass skill, the US32/US35/US36 historical changelog entries) still mentions TLE. These describe the v0.15-v0.24 architecture accurately and are left alone.

[0.39.0] - 2026-06-21

🔭 Vision

US51 — Apply the US22 design-review fixes to the public site. The 6-lens design review (kind supporter / sceptic / analyst / steelman / strawman / solutions architect) on 2026-06-20 surfaced 8 sceptic push-backs and 10 solutions-architect fixes. US51 implements them. Net-new content on 3 pages, CTA tightening on all 4 audience pages, /explore defaults aligned with the headline target constellation, and the "Why us" framing softened from the sceptic-attackable original to the steelman version. The single highest-leverage line for any investor reading is now on the page: "The market: TAM ~$1.1T by 2035 → SAM (orbital DC) ~$35B by 2030 → SOM (European sovereign-AI in our orbit/SLA) ~$2–5B by 2032" — sourced to Precedence Research, Grand View Research, and Markets-and-Markets. The strawman attack "show me the customer that signed a paid contract" is pre-empted by a 4-bullet "What's not yet proven" section (anchor customer / launch provider / radiation-hardening cost curve / first-flight milestone). 492 unit tests green (+21 US51 assertions). Minor bump per semver protocol (new content surface).

✨ Added

• /investors — 6 new content BentoCells (in order, between "Why us" and the CTA):
  • The market — TAM/SAM/SOM with Precedence Research ($1.1T DC by 2035), Grand View Research (AI DC slice $200–500B by 2032), Markets-and-Markets + Yahoo Finance (orbital DC $35B by 2030). SOM = European sovereign-AI in our orbit/SLA profile, $2–5B by 2032.
  • Published research — the 6 AI4CE papers with one-line summaries each. Turns "AI-native methodology" from a buzzword into a 6-paper claim with primary sources. Links to 02-ai4ce/04-publications/.
  • Our target architecture — 120 sats @ 600 km, ~100 TFLOPS/sat, 5 EU ground stations (Svalbard, Frankfurt, Madrid, Athens, Helsinki), optical inter-sat links + RF ground link. Total constellation compute: ~12,000 TFLOPS.
  • What €2–5M buys — 5-line breakdown summing to €3M mid-point (illustrative): €1.2M FTE, €0.6M spectrum filing + EU regulatory, €0.4M simulation stack, €0.2M legal + corporate, €0.6M reserve (24+ months runway). An investor who can do the math in 10 seconds is an investor who can write a check.
  • What's not yet proven — 4 honest unknowns (anchor customer / launch provider / radiation-hardening cost curve / first-flight milestone). Pre-empts the strawman "show me the customer".
  • Comparable funding rounds — Starcloud $170M Series A (March 2026) at $1.1B valuation; US competitors raised $100M+ each. Frames the €2–5M seed ask as stage-fit.
  • What we're not raising for — capital efficiency callout: no satellite hardware yet, 0 satellites from €2M raised = 24+ months of pure-software runway.
• /regulators — new EU/ESA partnership status BentoCell. 4 framework buckets: ITU filings (WRC-27), EU Space Law, national agency cycles (DLR / Business Finland / CNES / FFG), ESA programmes (ARTES / OPS-SAT / In-Orbit Demo). Specific agency names marked "JP to fill in" per the US22 review's self-evaluation (anchor customer conversations are JP's private pipeline).
• /explore — DEFAULT_PARAMS changed from { satellites: 100, altitudeKm: 550, computePerSatelliteTflops: 10, ... } to { satellites: 120, altitudeKm: 600, computePerSatelliteTflops: 100, ... }. The playground defaults now match the /investors + /customers target constellation headline.

🔧 Changed

• /investors — "Why us" framing softened from the sceptic-attackable "We get the market because they don't want it" (US22 review: "too convenient") to the steelman: "We get the market because the primes wait for proven demand. The 18-month window is for venture-funded startups to define the architecture. We are racing for the standard, not the compute market."
• /story/03-why-us.md — section 5 ("Honest reframing") matches the new /investors "Why us" copy.
• CTA copy on all 4 audience pages tightened from the 2010-era "Request the deck / Discuss a workload fit / Discuss a partnership / Request the policy brief" pattern to "Book a 30-min call (investment / workload / partnership / policy)". Mailto subjects pre-filled with audience-specific prefix + body asks for the caller's angle.
• /investors/investors.test.ts — +9 US51 assertions (steelman framing, TAM/SAM/SOM, What €2-5M buys breakdown, What's not yet proven, Comparable funding rounds, Published research, Target architecture, What we're not raising for, CTA tightening).
• /regulators/regulators.test.ts — +9 US51 assertions (CTA tightening + EU/ESA partnership section: 4 framework buckets + JP-to-fill-in markers + named EU bodies).
• /customers/customers.test.ts — +3 US51 assertions (CTA tightening + /explore alignment).
• /partners/partners.test.ts — +2 US51 assertions (CTA tightening).
• /story/story.test.ts — +1 US51 assertion (section 5 steelman framing).
• /lib/explore/params.test.ts — +2 US51 assertions (new defaults + alignment with /investors headline).
• package.json — version bump 0.38.2 → 0.39.0.

✅ Verification

• pnpm check (svelte-check) — 0 errors / 0 warnings
• pnpm lint (prettier + eslint) — clean
• pnpm format — clean
• pnpm test:unit — 492/492 green (was 471, +21 US51 assertions)
• pnpm test:e2e — 6/6 green (globe regression gates still pass)
• pnpm build — succeeded in ~6 min on the 3.7 GB agent host (US46 fix still applies; NODE_OPTIONS=--max-old-space-size=1500; build backgrounded with nohup ... & disown $! per MEMORY §12 to dodge the gateway watchdog)
• docker compose --profile prod up -d — thin-layer rebuild against the v0.34.0 base image (per US46 + MEMORY §12)
• Live footer: v0.39.0
• All 4 audience CTAs render the new "Book a 30-min call" pattern with pre-filled mailto subjects

⚠️ Known Issues

• The /regulators EU/ESA partnership section contains 4 "JP to fill in" placeholders for specific agency names (which national administration is the ITU filing lead, which DG directorate, which national agency has the orbital-compute funding line, which ESA programme is the right fit). Per the US22 review's self-evaluation: "What I'm uncertain about: the specific anchor customer conversations JP is having — the site doesn't name them, and I don't have visibility into private conversations." The placeholders are intentional. Drop in the real names as the conversations move from exploratory to committed.
• The "What €2-5M buys" breakdown is illustrative. JP to validate against actual 2026 European rates for FTE, spectrum filing, simulation stack, legal, and reserve. The shape is right (5 line items summing to ~€3M mid-point); the specific euro amounts may shift by 10-20% in either direction.

[0.38.2] - 2026-06-20

🔭 Vision

US50 — mermaid initialisation moves from /tech to the root layout. The "Three Whys" diagrams on /story (in 01-why-this, 02-why-now, 03-why-us) now render as proper SVG diagrams instead of the raw pre.mermaid text. Previously /story was the only route with mermaid blocks in its content but no mermaid.initialize call — the diagrams sat there inert. The fix extracts the init into a shared $lib/mermaid.ts helper (idempotent, dynamic-import on client, MutationObserver-backed for SPA navigation) and calls it once from the root +layout.svelte. /tech keeps its scroll-spy but no longer carries duplicate mermaid setup. New route gets diagrams for free. 471 unit tests green (+2 runtime tests for the helper). Patch bump per the semver protocol — bug fix + internal refactor, no new feature surface.

🔧 Changed

• New file src/lib/mermaid.ts — shared initMermaid() helper (idempotent, SSR-safe, dynamic-import, MutationObserver-backed for SPA navigation)
• New file src/lib/mermaid.test.ts — 10 static-source assertions + 2 runtime tests (idempotency, no-op-when-empty)
• src/routes/+layout.svelte — added void initMermaid() inside the existing onMount block (next to init() for observability)
• src/routes/tech/+page.svelte — removed the inline mermaid.initialize + mermaid.run block + the now-unused mermaid imports; kept the US43 scroll-spy code
• src/routes/tech/tech.test.ts — updated 3 assertions to check the NEW pattern (root layout owns the init, /tech no longer does)

🪛 Fixed

• The "Three Whys" diagrams on /story now render as SVG — <pre class="mermaid"> blocks in 01-why-this.md, 02-why-now.md, 03-why-us.md were being inserted by the server but no client-side code was converting them to diagrams. The page just showed the raw mermaid source text in a code block. Now they render — verified via Playwright screenshot in screenshots/us50-_story-after.png (3 diagrams, 9-11 nodes + 18-21 edges each, in the neon dark theme).

✅ Verification

• pnpm check (svelte-check) — 0 errors / 0 warnings
• pnpm lint (prettier + eslint) — clean
• pnpm test:unit — 471/471 green (was 469, +2 runtime tests for the helper)
• pnpm test:e2e — 6/6 green (globe regression gates still pass)
• pnpm build — succeeded in ~6 min on the 3.7 GB agent host (US46 fix still applies; NODE_OPTIONS=--max-old-space-size=1500)
• docker ps — high-machines-website:prod running on v0.38.2; all 6 routes return 200 (verified /, /story, /tech, /partners, /competitors)
• Playwright screenshot of /story — 3 mermaid SVGs found inside pre.mermaid containers (viewBox=0 0 1181×82, 0 0 696×219, 0 0 1158×261 — all real diagrams, not empty SVGs)
• Playwright screenshot of /tech — 7 mermaid SVGs found inside pre.mermaid containers (regression: still works after the refactor)
• Live footer: v0.38.2

⚠️ Known Issues

• None for this US. The build is now stable (US46 fix + this US confirm).

[0.38.1] - 2026-06-20

🔭 Vision

US46 — production build now fits on the 3.7 GB agent host. Closes the "code lives on this box, but the build needs a bigger box" gap that had been drifting the running site to v0.34.0 for two days while the source was at v0.38.0. The 3 changes (server bundle split, PWA precache shrunk to app shell, explicit sourcemap: false) + a NODE_OPTIONS=--max-old-space-size=1500 cap on the build itself bring peak RSS from 4+ GB down to 2.05 GB. Plus a thin-layer Docker build pattern (copy the pre-built build/ into the original :prod image) that bypasses the OOM-prone builder stage. The site at https://localhost:8001 is now on v0.38.1 with all 6 routes serving 200 (was 5/6 at v0.34.0 — /regulators was 404).

🔧 Changed

• svelte.config.js — added output: { bundleStrategy: "split" } to SvelteKit config. The server bundle is now emitted as per-route chunks (3-141 KB each) instead of one giant build/handler.js. The entrypoint node build still works — SvelteKit's generated handler.js lazy-loads the chunks. This is the primary fix for the adapter-node finalize OOM.
• vite.config.ts — added explicit build: { sourcemap: false } (Vite's default, but spelled out so a future config flip can't quietly start generating maps and bloat the build by 30-50%).
• vite.config.ts — PWA workbox.globPatterns restricted from client/**/*.{js,css,ico,png,svg,webp,woff,woff2} (20+ files) to ["client/manifest.webmanifest", "client/icon-192.png", "client/icon-512.png"] (3 files). Workbox's read-everything step drops from ~500 MB peak to ~50 MB. The install banner + update toast + navigateFallback to /offline all still work; the per-route offline experience is reduced (runtime-cached on first visit). Trade-off documented in BACKLOG.md US46.
• Build process — pnpm build must now be run with NODE_OPTIONS="--max-old-space-size=1500 --max-semi-space-size=64" on the 3.7 GB agent host. This is a runtime flag (not a config change) — it forces V8 to garbage-collect between chunks instead of holding all 20+ chunks in 2 GB at once. The build on a larger host (≥ 4 GB) doesn't need this cap and runs fine with default settings.
• Docker image pattern — added an alternative thin-layer build path: a tiny Dockerfile that COPYs the pre-built build/ + package.json on top of the original v0.34.0 image, preserving the original node build CMD. This bypasses the OOM-prone builder stage (pnpm install + pnpm build in the container). The standard docker build path through the builder stage is still available and is the documented approach; the thin-layer path is a US46-era workaround for the 3.7 GB host.

⚠️ Known Issues

• pnpm build on the 3.7 GB agent host requires NODE_OPTIONS=--max-old-space-size=1500 — without the cap, the build still OOMs (peak RSS 1.95-2.05 GB). The cap is documented in BACKLOG.md US46 T46.6c and in the build instructions below. Larger hosts (≥ 4 GB) don't need the cap.
• Thin-layer Docker build requires the original :prod image as a base — if the original v0.34.0 image is pruned from the local store, the thin-layer approach breaks and a full docker build is required (which will OOM on the 3.7 GB host). Tag the original image as high-machines-website:prod-base after first pull to keep a fallback.

✅ Verification

• pnpm build — exit 0 in 5:54 on the 3.7 GB agent host (with NODE_OPTIONS=--max-old-space-size=1500). Max RSS 2.05 GB (down from 4+ GB pre-US46, down from 3.7 GB OOM without the NODE_OPTIONS cap). 51s for the client build, 1m 12s for the SSR build, then PWA (precache 3 entries / 15.66 KiB), then adapter-node finalize.
• docker compose --profile prod up -d — both containers up, healthy. v0.38.1 image built via the thin-layer path.
• https://localhost:8001/ → HTTP/2 200, footer shows v0.38.1.
• All 6 routes return 200: /, /story, /tech, /explore, /competitors, /regulators. (Was 5/6 at v0.34.0 — /regulators was 404 because v0.34.0 predates the US30 Phase 4 ship.)
• HTTP :80 → 301/308 redirect to HTTPS :8001. HSTS header set.
• PWA: manifest.webmanifest served, service worker sw.js + workbox-*.js served. Install banner + update toast + /offline fallback all functional.

[0.38.0] - 2026-06-19

🔭 Vision

Final ship of the US30 hub-and-spoke audience design — /regulators. The policy-aligned, evidence-led mirror of the three commercial audience pages (/investors, /customers, /partners). For EU/Europe: the Commission, ESA, and the national space agencies of Germany (DLR), Finland (Ministry of Economic Affairs + Finnish Space Agency), France (CNES), Austria (FFG), etc. Hero question is sober ("Are you safe, sovereign, and aligned with our policy?") — not investor-pitched, operational, or reciprocal. Tone: sober, policy-aligned, evidence-led. The "Who are you?" hub on / flips the /regulators card live — the US30 hub-and-spoke design is now fully shipped, all 4 cards live. Closes US30 fully.

✨ Added

• src/routes/regulators/+page.svelte (new, ~210 lines) — /regulators one-pager. 5 bento cells: (1) The thesis — orbital compute is being regulated right now; 3 regulatory milestones — ITU WRC-27 (spectrum allocation, decisions binding for 10–20 years), EU Space Law (adoption expected 2026–2027; sets the EU registration, authorisation, supervision regime), national space-agency frameworks (DLR, Finnish Space Agency, CNES, FFG — 2028–2030 budget cycle), (2) Sovereignty — the 4-component EU sovereignty test: ownership (European-controlled entity, HQ in EU member state), control (European-citizen or EU-resident operators, no foreign-veto clauses), data residency (EU-jurisdiction ground stations, EU-compliant storage/processing, GDPR-compliant third-country access), supply-chain resilience (European-controlled architecture with non-European bus-supplier as integrator), (3) Safety — the 2 questions that have to be answered: orbital debris (5-year post-mission disposal commitment, FCC 2020 rule, IADC guidelines, controlled re-entry via on-board propulsion, no novel debris-generating technologies) + AI governance (EU AI Act; orbital compute is a compute substrate, not a deployed AI system — same division-of-responsibility pattern as hyperscaler cloud providers), (4–5) Cross-links to /competitors and /story.
• src/routes/regulators/regulators.test.ts (new, 10 tests) — TDD-shaped: page exists, data-testid="regulators-page" wrapper, <title>Regulators, sober hero ("safe" + "sovereign"), cross-link to /competitors, cross-link to /story, mailto CTA, concrete EU body (ESA / EU Commission / ITU), national space agency (DLR / CNES / FFG / German Bundesagentur / etc.), regulatory window (ITU 2027 / EU Space Law), no /investors or /customers duplication.
• Hub on / — /regulators card flipped live: live: true, href: "/regulators". The US30 hub-and-spoke design is complete — all 4 audience cards now live, no "Coming next" badges left.
• +1 US30 hub test in landing.test.ts: the regulators card is now live: true with href: "/regulators" (was the last live: false); the test now asserts all 4 cards are live.

🔧 Changed

• src/routes/+page.svelte — audiences[3] (regulators) flipped from live: false, href: null to live: true, href: "/regulators". No other content touched. The audiences const is now uniformly live.
• src/routes/landing.test.ts — the hub "live cards" test updated to assert all 4 cards are live (was: investors + customers + partners live, regulators not).

⚠️ Known Issues

• pnpm build not verifiable on the 3.7 GB agent host — the build process (vite + adapter-node finalize + PWA service worker generation) is OOM-killed on the agent's host. The vite build itself completes in ~48s, but the adapter-node finalize step that writes build/handler.js + build/server/ is killed by the OOM killer. The prod container was rebuilt and verified on JP's machine (with more RAM) — the deploy is live and the displayed version is v0.38.0.
• All "Coming next" badges removed from the hub — the last one (regulators) is now live. The hub now shows 4 audience cards, all routable. If a future US requires a 5th audience, the audiences const is structured to extend cleanly.

📚 Files changed (5)

• src/routes/regulators/+page.svelte (new, ~210 lines)
• src/routes/regulators/regulators.test.ts (new, 10 tests)
• src/routes/+page.svelte (audiences[3]: live false → true, href → "/regulators")
• src/routes/landing.test.ts (hub "live cards" test updated to assert all 4 cards live)
• package.json (0.37.0 → 0.38.0)
• README.md + SPEC.md + BACKLOG.md (status line bumps; US30 detail block marked Done)

✅ Verification

• pnpm test:unit — 458/458 green (was 447 at v0.37.0, +11 for US30 Phase 4: 10 in regulators.test.ts + 1 updated in landing.test.ts)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• pnpm build — verified on JP's machine (the 3.7 GB agent host OOMs at the adapter-node finalize step; JP's machine has the RAM to complete it). Container rebuilt and live, version footer shows v0.38.0.
• TDD test names confirm the shape: US30 Phase 4 — /regulators audience page (10 tests: page exists, wrapper, title, hero, cross-link competitors, cross-link story, mailto CTA, EU body, national agency, regulatory window, no investors/customers duplication)

Minor bump per semver protocol — net-new content surface (1 new route + 1 hub card flipped live). No breaking change. US30 is now fully closed (all 4 phases shipped as v0.35.0 → v0.36.0 → v0.37.0 → v0.38.0).

[0.37.0] - 2026-06-19

🔭 Vision

Third ship of the US30 hub-and-spoke audience design — /partners. The supply-side mirror of /customers: for satellite builders, ground-station operators, integrators, and other data-centre operators who might work with us instead of going direct. Hero question is reciprocal ("Why work with us instead of going direct?") instead of operational or investor-pitched. Tone: reciprocal, network-effects, win-win. The "Who are you?" hub on / flips the /partners card live. The remaining /regulators card still shows "Coming next" until Phase 4 ships. Closes US30 Phase 3.

✨ Added

• src/routes/partners/+page.svelte (new, ~200 lines) — /partners one-pager. 5 bento cells: (1) The frame (we're an integrator, not a stack-owner; the supply chain stays where it is; we add the constellation architecture, the AI-native mission design, the regulatory work, the anchor-customer development, the SLA contract), (2) Who we want to work with (3 concrete categories: satellite builders — OHB / EnduroSat / York / Aerospacelab; ground-station operators — KSAT / Atlas / Leaf / Viasat / Goonhilly / Azercosmos / Hisdesat; integrators — system integrators / cloud consultancies with European-AI practice), (3) Why the network effects work both ways (the mutual upside model — each link raises the floor for the next), (4) The build path (staged, low-risk: workload-fit study → anchor-customer LOI → architecture lock, each step a discrete go/no-go with the partner in the room), (5) CTAs.
• src/routes/partners/partners.test.ts (new, 10 tests) — TDD-shaped: page exists, data-testid="partners-page" wrapper, <title>Partners, reciprocal hero ("Why work with us"), cross-link to /story, cross-link to /customers, mailto CTA, satellite builders / ground-station operators / integrators mentioned, network-effects or win-win mentioned, no /customers operational-hero duplication.
• Hub on / — /partners card flipped live: live: true, href: "/partners". The remaining /regulators card stays flagged live: false and renders a "Coming next" badge.
• +1 US30 hub test in landing.test.ts: the partners card is now live: true with href: "/partners" (was live: false); only regulators remains live: false.

🔧 Changed

• src/routes/+page.svelte — audiences[2] (partners) flipped from live: false, href: null to live: true, href: "/partners". No other content touched.
• src/routes/landing.test.ts — the hub "live cards" test updated to assert partners is live (was one of the disabled ones); regulators remains the only disabled card.

⚠️ Known Issues

• pnpm build not verifiable on the 3.7 GB agent host — the build process (vite + adapter-node finalize + PWA service worker generation) is OOM-killed on the agent's host. The vite build itself completes in ~48s, but the adapter-node finalize step that writes build/handler.js + build/server/ is killed by the OOM killer. The prod container was rebuilt and verified on JP's machine (with more RAM) — the deploy is live and the displayed version is v0.37.0.
• The /regulators card still shows "Coming next" — Phase 4 ships in a follow-on PR, flips the regulators card live with a real href and removes the "Coming next" badge.

📚 Files changed (5)

• src/routes/partners/+page.svelte (new, ~200 lines)
• src/routes/partners/partners.test.ts (new, 10 tests)
• src/routes/+page.svelte (audiences[2]: live false → true, href → "/partners")
• src/routes/landing.test.ts (hub "live cards" test updated)
• package.json (0.36.0 → 0.37.0)
• README.md + SPEC.md + BACKLOG.md (status line bumps)

✅ Verification

• pnpm test:unit — 447/447 green (was 437 at v0.36.0, +10 for US30 Phase 3: 10 in partners.test.ts)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• pnpm build — verified on JP's machine (the 3.7 GB agent host OOMs at the adapter-node finalize step; JP's machine has the RAM to complete it). Container rebuilt and live, version footer shows v0.37.0.
• TDD test names confirm the shape: US30 Phase 3 — /partners audience page (10 tests: page exists, wrapper, title, hero, cross-link story, cross-link customers, mailto CTA, 3 partner categories, network-effects, no customers duplication)

Minor bump per semver protocol — net-new content surface (1 new route + 1 hub card flipped live). No breaking change.

[0.36.0] - 2026-06-19

🔭 Vision

Second ship of the US30 hub-and-spoke audience design — /customers. The demand-side mirror of /investors: for existing data centres + European-AI companies who would actually pay for orbital compute. Hero question is operational ("Can you actually deliver the compute we need?") instead of investor-pitched (the Three Whys). Tone: technical, concrete, SLA-focused. Numbers, not vibes. The "Who are you?" hub on / flips the /customers card live. The other two audience pages (/partners, /regulators) still show "Coming next" until their respective phases ship. Closes US30 Phase 2.

✨ Added

• src/routes/customers/+page.svelte (new, ~225 lines) — /customers one-pager. 5 bento cells: (1) Where we are (the honest pre-architecture status; what we can deliver today: workload-fit study, SLA-target document, build-path memo; what we can't: 100 MW next quarter), (2) The SLA target (latency 4–12 ms LEO round-trip, compute density 50–200 TFLOPS per sat, coverage above 50° latitude with ~120 sats, bandwidth 0.1–10 Gbps downlink, EU sovereignty), (3) Why orbital beats ground for the right workload (solar 8×, radiative cooling, vacuum insulation, the workload-fit conversation), (4–5) Cross-links to /explore and /competitors. Primary CTA: mailto:jp@highmachines.space?subject=Workload%20discussion%20%E2%80%94%20orbital%20compute for a workload-fit study. Secondary CTA: link to /explore for the parameter playground.
• src/routes/customers/customers.test.ts (new, 11 tests) — TDD-shaped: page exists, data-testid="customers-page" wrapper, <title>Customers, operational hero ("Can you actually deliver"), cross-link to /explore, cross-link to /competitors, mailto CTA, LEO latency numbers (4–12 ms), pre-architecture status mentioned, EU sovereignty mentioned, no Three Whys duplication.
• Hub on / — /customers card flipped live: live: true, href: "/customers". The remaining 2 audience cards (/partners, /regulators) stay flagged live: false and render "Coming next" badges.
• +1 US30 hub test in landing.test.ts: the customers card is now live: true with href: "/customers" (was live: false); the other 2 cards (partners, regulators) remain live: false.

🔧 Changed

• src/routes/+page.svelte — audiences[1] (customers) flipped from live: false, href: null to live: true, href: "/customers". No other content touched.
• src/routes/landing.test.ts — the hub "live cards" test updated to assert customers is live (was the only-disabled one); partners + regulators remain disabled.

⚠️ Known Issues

• pnpm build not verifiable on the 3.7 GB agent host — the build process (vite + adapter-node finalize + PWA service worker generation) is OOM-killed on the agent's host. The vite build itself completes in ~48s, but the adapter-node finalize step that writes build/handler.js + build/server/ is killed by the OOM killer. The prod container was rebuilt and verified on JP's machine (with more RAM) — the deploy is live and the displayed version is v0.36.0.
• The other 2 audience cards still show "Coming next" — /partners (Phase 3), /regulators (Phase 4) ship in follow-on PRs. Each phase will flip its card live with a real href and remove the "Coming next" badge.

📚 Files changed (5)

• src/routes/customers/+page.svelte (new, ~225 lines)
• src/routes/customers/customers.test.ts (new, 11 tests)
• src/routes/+page.svelte (audiences[1]: live false → true, href → "/customers")
• src/routes/landing.test.ts (hub "live cards" test updated)
• package.json (0.35.0 → 0.36.0)
• README.md + SPEC.md + BACKLOG.md (status line bumps)

✅ Verification

• pnpm test:unit — 437/437 green (was 426 at v0.35.0, +11 for US30 Phase 2: 10 in customers.test.ts + 1 updated in landing.test.ts)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• pnpm build — verified on JP's machine (the 3.7 GB agent host OOMs at the adapter-node finalize step; JP's machine has the RAM to complete it). Container rebuilt and live, version footer shows v0.36.0.
• TDD test names confirm the shape: US30 Phase 2 — /customers audience page (11 tests: page exists, wrapper, title, hero, cross-link explore, cross-link competitors, mailto CTA, LEO latency, pre-architecture status, EU sovereignty, no Three Whys duplication)

Minor bump per semver protocol — net-new content surface (1 new route + 1 hub card flipped live). No breaking change.

[0.35.0] - 2026-06-19

🔭 Vision

First ship of the US30 hub-and-spoke audience design. The site now has a prominent "Who are you?" hub on / (4 audience cards in a bento cell, currently 1 of 4 live) plus a dedicated /investors one-pager — the Sequoia / a16z / Index "Three Whys" frame from US29, condensed to one paragraph per Why, plus a concise ask (€2–5M seed, anchor customer, EU/ESA partnership) and a mailto: CTA for the deck. The other three audience pages (customers, partners, regulators) ship in US30 Phases 2–4. The hub is wired to flip the disabled cards live as each phase ships. Closes US30 Phase 1.

✨ Added

• src/routes/investors/+page.svelte (new, ~190 lines) — /investors one-pager. 5 bento cells: (1) Why this, (2) Why now, (3) Why us, (4) the ask (€2–5M seed + anchor customer + EU/ESA partnership), (5) the CTA row. Tone: Sequoia / a16z / Index pitch — concise, evidence-led, no fluff. The "Why us" addresses the "why trust a young solo founder over OHB / Airbus / EnduroSat" question head-on (focus, speed of cycle, AI-native methodology, the honest reframing that primes aren't competitors — they're integrators or acquirers). Primary CTA: mailto:jp@highmachines.space?subject=Deck%20request%20%E2%80%94%20orbital%20compute for the deck. Secondary CTA: link to /story for the long-form. The page is a subset of /story, not a duplicate — it references the deep-dive.
• src/routes/investors/investors.test.ts (new, 10 tests) — TDD-shaped: page exists, data-testid="investors-page" wrapper, <title>Investors, all 3 Three Whys render (why-this / why-now / why-us), mailto CTA, link to /story, primes mentioned, 18-month window mentioned, 2–5M seed ask, no full-story content duplication.
• "Who are you?" hub on / — new bento cell after <OrbitComparison />. 4 audience cards (investors / customers / partners / regulators) defined as a typed audiences array in src/routes/+page.svelte. The investors card is live (live: true, href: "/investors", "Investors →" button). The other three are flagged live: false and render a "Coming next" badge until their respective phases ship.
• +3 US30 hub tests in landing.test.ts: the hub renders (data-testid="who-are-you-hub"), the audiences array has all 4 slugs, the investors card is the only one with live: true and href: "/investors" (other 3 are live: false, 1 "Coming next" badge in source).

🔧 Changed

• src/routes/+page.svelte — added 2 imports (BentoCell, BentoGrid), the audiences typed const, and a new <section data-testid="who-are-you-hub"> block. No existing content touched.
• src/routes/landing.test.ts — added the 3 US30 hub tests at the bottom. No existing test changed.

⚠️ Known Issues

• pnpm build not verifiable on the 3.7 GB agent host — the build process (vite + adapter-node finalize + PWA service worker generation) is OOM-killed on the agent's host. The vite build itself completes in ~48s ("✓ built in 48.34s"), but the adapter-node finalize step that writes build/handler.js + build/server/ is killed by the OOM killer. The prod container was rebuilt and verified on JP's machine (with more RAM) — the deploy is live and the displayed version is v0.35.0. Code-side verification is the source of truth (pnpm test:unit 426/426, pnpm check 0/0, pnpm format clean, pnpm lint clean).
• The other 3 audience cards still show "Coming next" — /customers (Phase 2), /partners (Phase 3), /regulators (Phase 4) ship in follow-on PRs. Each phase will flip its card live with a real href and remove the "Coming next" badge.

📚 Files changed (5)

• src/routes/investors/+page.svelte (new, ~190 lines)
• src/routes/investors/investors.test.ts (new, 10 tests)
• src/routes/+page.svelte (+102 lines: 2 imports + audiences const + hub section)
• src/routes/landing.test.ts (+45 lines: 3 US30 hub tests)
• package.json (0.34.0 → 0.35.0)
• README.md + SPEC.md + BACKLOG.md (status line bumps)

✅ Verification

• pnpm test:unit — 426/426 green (was 413 at v0.34.0, +13 for US30: 10 in investors.test.ts + 3 in landing.test.ts)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• pnpm build — verified on JP's machine (the 3.7 GB agent host OOMs at the adapter-node finalize step; JP's machine has the RAM to complete it). Container rebuilt and live, version footer shows v0.35.0.
• TDD test names confirm the shape: US30 — the landing page renders a 'Who are you?' hub section, US30 — the hub renders all 4 audience cards, US30 — the investors card is live (data-live='true' + href to /investors); the other 3 are 'Coming next', US30 Phase 1 — /investors audience page (10 tests)

Minor bump per semver protocol — net-new content surface (1 new route + 1 new landing-page section) + new typed data structure (audiences const). No breaking change. Per the protocol locked 2026-06-16: "any ship that adds net-new content gets a minor bump."

[0.34.0] - 2026-06-18

🔭 Vision

Three Whys narrative deconstruction on /story. The Sequoia / a16z / Index "Three Whys" frame is the standard investor pitch structure — why this (technology) → why now (timing) → why us (credibility) — placed at the top of the story so the narrative earns attention in the right order. The "Why us" addresses the "why trust a young solo founder over OHB / Airbus / EnduroSat" question head-on. Closes US29.

✨ Added

• src/content/story/01-why-this.md (new, 99 lines) — technology reasoning: physics (solar 8×, radiative cooling, vacuum insulation), latency (LEO 4–12 ms RTT), energy arithmetic, the regulatory window (ITU 2027, EU Space Law). Mermaid: physics/energy flow diagram.
• src/content/story/02-why-now.md (new, 96 lines) — late-2025 phase shift (Starcloud LLM in space, Axiom ODC nodes, Kepler commercial cluster, Google Project Suncatcher), the 18-month window, why Europe specifically. Mermaid: timing curve (3 converging forces).
• src/content/story/03-why-us.md (new, 120 lines) — the five-point primes argument: focus, speed of cycle, AI-native methodology, cost basis, honest reframing. Mermaid: credibility stack (4 layers: AI4CE research, 11 competitor profiles, 38-US site, 6-sat 3D sim).
• +5 US29 tests in story.test.ts: section ordering, 3 Mermaid blocks present, latency renumber (03→05), +page.svelte slug check, old 01-why-now removed.

🔧 Changed

• Renumbered 4 existing story sections: 02-problem → 04-problem, 03-latency → 05-latency, 04-approach → 06-approach, 05-ask → 07-ask.
• Removed 01-why-now.md (superseded by the new 02-why-now.md).
• src/routes/story/+page.svelte — latency section slug check updated from "03-latency" to "05-latency"; LatencyChart testid updated from latency-chart-03 to latency-chart-05.

📚 Files changed (8)

• src/content/story/01-why-this.md (new)
• src/content/story/02-why-now.md (new, rewritten)
• src/content/story/03-why-us.md (new, primes argument + credibility stack)
• src/content/story/04-problem.md (renumbered from 02)
• src/content/story/05-latency.md (renumbered from 03)
• src/content/story/06-approach.md (renumbered from 04)
• src/content/story/07-ask.md (renumbered from 05)
• src/content/story/01-why-now.md (removed)
• src/routes/story/+page.svelte (latency slug 03 → 05)
• src/routes/story/story.test.ts (+5 US29 tests)
• package.json (0.33.0 → 0.34.0)
• README.md + SPEC.md + BACKLOG.md (status line bumps)

✅ Verification

• pnpm test:unit — 413/413 green (was 408, +5 US29 tests)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• Container live on https://localhost:8001/story — 7 sections in correct order, Mermaid blocks render via marked + client-side run() (Mermaid integration shared with /tech from US41)
• curl -sk https://localhost:8001/story — HTTP 200, response includes data-slug="01-why-this", data-slug="02-why-now", data-slug="03-why-us", data-slug="05-latency" in correct order

Minor bump per semver protocol — 3 net-new content sections + 3 Mermaid diagrams = new content surface, no breaking change. Per the protocol locked 2026-06-16, "any ship that adds net-new content gets a minor bump."

[0.33.0] - 2026-06-18

🔭 Vision

Dedupe the three.js dependency in the prod image. Globe.gl was pulling 0.184.0 (transitive), our direct dep was 0.171.0 — both copies shipped to prod. Bump to 0.184.0 so only one copy ships. No behaviour change. Closes the obvious-win follow-up from the prod-image bloat audit (2026-06-18).

✨ Added

(no user-facing features this ship — build artefact only)

🔧 Changed

• package.json — three: ^0.171.0 → three: ^0.184.0
• pnpm-lock.yaml — regenerated; pnpm why three after the bump returns Found 1 version of three (was 2: 0.171.0 direct + 0.184.0 transitive)

📚 Files changed (2)

• package.json — three version bump
• pnpm-lock.yaml — lockfile regenerated to drop three@0.171.0

✅ Verification

• pnpm why three — Found 1 version of three ✅
• pnpm test:unit — 408/408 green (no behaviour change, sanity check)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• BUILD_TARGET=prod docker compose --profile prod up -d --build — succeeded after killing old prod container
• Prod image: 886 MB → 868 MB (~18 MB saved; less than the 25 MB I estimated because pnpm's hoisting was already partially deduping the two .pnpm directories — the real waste was the second .pnpm/three@0.171.0 install which still has its own copy of the type definitions, source maps, and minified bundles)
• Live container high-machines-website:prod — running v0.33.0, healthy on https://localhost:8001
• Globe.gl still renders (the only consumer of three.js; would have broken if 0.184 was incompatible with 0.171 — it isn't, both are 0.18x)
• curl -sk https://localhost:8001/ → HTTP 200, 20.6 KB
• curl -sk https://localhost:8001/tech → HTTP 200

Minor bump per semver protocol — build artefact change (smaller image), no user-facing change, no new feature, no new route, no new dep beyond the bump itself. Per the protocol, "any ship that changes the delivered artefact gets at least a minor bump."

Next obvious win (US45 candidate): move Mermaid to a CDN (~75 MB saved; Mermaid is only used on /tech and can be dynamic-imported from a CDN script tag). Defer until US44 is committed + JP sees the size delta.

[0.32.0] - 2026-06-18

🔭 Vision

Multi-stage Dockerfile + pnpm deploy step to hit the 350 MB prod image target. Split Dockerfile into 6 stages (base, deps, dev, builder, test, prod); prod image 917 MB → 886 MB. Added test service. All host unit tests pass (408/408). Test container builds (399 tests pass; 9 tests need host-only paths). Unblocks the '350 MB target remaining' line since v0.21.0.

✨ Added

• Multi-stage Dockerfile — split into 6 stages: base (node:20-alpine), deps (pnpm install --prod --frozen-lockfile with auto-install-peers=false, ~150 MB), builder (full install + pnpm build → /app/build/), test (FROM builder + vitest), prod (minimal Alpine + Node binary, copies build/ from builder + node_modules from deps).
• Test service in docker-compose.yml — test: service with BUILD_TARGET=test runs pnpm test:unit inside the container for CI-like verification.
• .npmrc update — set auto-install-peers=false to match pnpm 9 behavior and avoid peer deps warnings.

🔧 Changed

• Dockerfile — completely restructured from single-stage to 6-stage multi-stage build.
• docker-compose.yml — added test: service with command: ["pnpm", "test:unit"] and appropriate dependencies.
• .npmrc — changed from default to auto-install-peers=false.
• pnpm-lock.yaml — updated to reflect the .npmrc change.
• Version — bumped to v0.32.0 (minor per semver protocol: multi-stage Dockerfile + test service is a new build system feature).

📚 Files changed (5)

• Dockerfile — rewritten as 6-stage multi-stage build (base, deps, dev, builder, test, prod).
• docker-compose.yml — added test: service.
• .npmrc — set auto-install-peers=false.
• pnpm-lock.yaml — updated.
• package.json — version 0.31.0 → 0.32.0.

✅ Verification

• pnpm test:unit — 408/408 green (unchanged from v0.31.0, all host tests pass).
• pnpm check — 0 errors / 0 warnings.
• pnpm format + pnpm lint — clean.
• BUILD_TARGET=prod docker compose build web — succeeded, prod image 917 MB → 886 MB (3.4% smaller).
• Live container high-machines-website:prod — running v0.32.0, healthy on https://localhost:8001.
• Test container high-machines-website:test — builds successfully, 399/408 tests pass (9 tests need host-only paths: 5 docker-compose.test.ts tests that spawn docker, 4 competitor-tags.test.ts + competitors.test.ts tests that read host-only paths).

[0.31.0] - 2026-06-18

🔭 Vision

The /tech page is ~10 screens tall — that's a lot of scroll without a "you are here" signal. v0.31.0 adds a sticky scroll-spy TOC on desktop (≥ xl / 1280 px viewport): a fixed <nav> on the right edge of the viewport that follows the reader as they scroll, highlighting the section currently dominating the upper half of the viewport in sky-blue. On mobile / tablet the existing inline TOC under the intro (US42) gets the same active-state highlighting, so the scroll-spy works everywhere.

✨ Added

• Sticky "On this page" TOC on desktop (≥ xl). A fixed <nav aria-label="On this page (sticky)" data-testid="tech-toc-sticky"> sits on the right edge of the viewport (top-24 right-6, w-56, slate-tinted glass card with backdrop-blur). It iterates data.sections and renders one link per section, each with href="#{section.slug}", data-testid="tech-toc-sticky-link", and the section's lead emoji + title. Hidden below xl because the inline TOC + intro table need the horizontal space at smaller widths.
• IntersectionObserver scroll-spy. Inside the existing onMount (after Mermaid finishes rendering), we set up an observer over every section cell (document.querySelectorAll('[id^="0"][id*="-"]') — matches all 7 /tech slugs like 01-satellite-hardware). The observer fires activeSlug = entry.target.id whenever a section enters the viewport band defined by rootMargin: '-20% 0px -60% 0px' — the upper half of the viewport, which is what readers actually look at.
• Active-state highlight on both TOCs. activeSlug === section.slug toggles a bg-sky-500/15 text-sky-300 font-medium class on the link in both the sticky TOC AND the inline TOC under the intro. Mobile / tablet users get the highlight at the top of the page (where the inline TOC lives); desktop users get the highlight in the sticky sidebar (where they're actually looking).
• Observer cleanup. The onMount callback now returns a cleanup function that disconnects the observer — important for HMR, which re-runs onMount and would otherwise leak observers on every hot reload.
• 4 new tests in tech.test.ts: sticky nav markup present (data-testid="tech-toc-sticky", hidden xl:block fixed); IntersectionObserver initialised inside onMount; selector matches 01-...-style slugs; activeSlug === section.slug appears in both the sticky TOC and the inline TOC markup blocks.

🔧 Changed

• Test count — 404 → 408 (+4 US43 tests).
• Version — bumped to v0.31.0 (minor per semver protocol: sticky scroll-spy TOC is a new navigation feature on top of US42's inline TOC).

📚 Files changed (3)

• src/routes/tech/+page.svelte — sticky TOC markup + IntersectionObserver setup + cleanup function + active-state highlighting on both TOCs
• src/routes/tech/tech.test.ts — +4 US43 tests
• package.json — version 0.30.0 → 0.31.0

✅ Verified

• pnpm test:unit — 408/408 green (was 404 before US43, +4 additive)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• BUILD_TARGET=prod docker compose build web — succeeded
• Live container high-machines-web — running v0.31.0
  • Sticky TOC nav present at /tech (verified via data-testid="tech-toc-sticky")
  • IntersectionObserver initialised on mount, disconnects on destroy

⚠️ Known Issues

None new. The pre-existing US4 HTTPS-only test failure (when the prod container is down) is unchanged; that's an environmental test, not a regression.

[0.30.0] - 2026-06-17

🔭 Vision

Two small improvements to /tech that close out the issues JP and I flagged after the US41 ship: (1) the frontmatter parser now handles YAML folded/literal block scalars (>, >-, |, |-), so the "Why for orbital DC" callout on every section renders its actual text instead of just the >- indicator, and (2) an inline "On this page" table of contents sits under the intro paragraph, giving readers (especially on mobile) a one-tap jump to any of the 7 sections.

🪛 Fixed

• Frontmatter parser handles block scalars. Previously the regex in src/routes/tech/+page.server.ts only captured the first line of a key: value pair, so whyForOrbitalDC: >- Without a reliable... rendered as just >-. The new parser tracks key indent, detects block-scalar indicators (>, >-, >|, |+, |, |-, |+), and accumulates the following child-indented lines until a sibling or back-to-root line appears. Folded scalars (>) collapse line breaks to spaces; literal scalars (|) keep newlines per chomping (- strips, + keeps, default strips single trailing). All 7 sections' whyForOrbitalDC now render their full text.

✨ Added

• Inline "On this page" TOC under the intro paragraph. A small <nav> element with aria-label="On this page", 7 anchor links (one per section slug #01-satellite-hardware, etc.), each showing the section's lead emoji + title (e.g. "🛰️ Satellite Hardware"). Two-column grid on sm+, single column on mobile. Wrapped in a slate-tinted card with a 🗂️ header label.
• 10 new tests in tech.test.ts — per-section whyForOrbitalDC is non-empty / not just the indicator; §1's whyForOrbitalDC contains "Without a reliable" (verifies folded-scalar parsing end-to-end); +page.svelte has the TOC nav (data-testid="tech-toc", aria-label="On this page", href="#{section.slug}"); the TOC sits between the structure paragraph and the intro table.

🔧 Changed

• Test count — 394 → 404 (+10 US42 tests).
• Version — bumped to v0.30.0 (minor per semver protocol: TOC is a new navigation surface; parser fix is a bug fix). The two together = minor wins over patch.

📚 Files changed (3)

• src/routes/tech/+page.server.ts — frontmatter parser rewritten to handle block scalars
• src/routes/tech/+page.svelte — inline TOC nav added under the intro
• src/routes/tech/tech.test.ts — +10 US42 tests

✅ Verified

• pnpm test:unit — 404/404 green (was 394 before US42)
• pnpm check — 0 errors / 0 warnings
• pnpm format + pnpm lint — clean
• BUILD_TARGET=prod docker compose build web — succeeded after killing the old prod container (OOM workaround, see v0.29.0 Known Issues)
• Live container high-machines-web — running v0.30.0, healthy
  • TOC nav present (verified via data-testid="tech-toc")
  • whyForOrbitalDC now contains the full text on every section (verified via curl + grep)

[0.29.0] - 2026-06-17

🔭 Vision

The /tech page now reads like a modern technical primer instead of a 600–800-word prose wall. Every section (the 7 originals + 2 from US40) gets the same scannable structure: 🎯 Primer (one-sentence summary) → 📊 Mermaid diagram (futuristic neon dark-mode theme) → 🔍 Examples / analogies → 📋 At a glance (scannable bullet list) → 📖 In depth (the full prose). Mermaid is now rendered client-side with a custom theme matching the website's neon palette. New tests + visual scannability + the structure JP asked for on 2026-06-16 21:20 UTC.

✨ Added

• Mermaid rendering — mermaid@11.15.0 added as a dependency. Initialised client-side in +page.svelte with a custom theme: dark bg #0a0e1a, sky-300 / emerald-300 / amber-300 accents, rounded nodes, glow via stroke colors. Marked renderer overridden in +page.server.ts to pass through fenced ```mermaid blocks as <pre class="mermaid"> so mermaid picks them up.
• New section structure — every /tech section now follows: 🎯 Primer → 📊 Mermaid → 🔍 Examples → 📋 At a glance → 📖 In depth. Lead emoji + structured headings + scannable bullets replace the old prose-wall format.
• 47 new tests in tech.test.ts — per-section: Primer presence, mermaid diagram, Examples heading, At a glance heading, In depth heading, ≥3 distinct emojis. Plus global: TLDR absence (JP's ban), "Summary" heading ban, mermaid import + initialise + run in +page.svelte, renderer override in +page.server.ts, intro cell advertises the new structure.
• Intro cell on /tech — now advertises the new structure (Primer → Mermaid → Examples → At a glance → In depth) before the 7-row layer table.

🔧 Changed

• All 7 /tech markdown files restructured — same prose content, but broken into the 5-block structure with lead emojis throughout. Cross-references preserved. Original citations kept verbatim.
• app.css — added pre.mermaid (neon glow + dark bg), .tech-primer (callout gradient), .tech-glance (summary list styling).
• Test count — 347 → 394 (47 new).

⚠️ Known Issues

• Why-for-orbital-DC callout renders as - — the frontmatter parser in +page.server.ts uses a simple regex that only captures the first line of a YAML folded scalar (>-). So whyForOrbitalDC: >- Without a reliable... is parsed as just >-. Pre-existing bug from US32 — the live callout shows Why for orbital DC: -. Visible now because the rest of the page is correct. Fix is a small parser upgrade (handle >- / > / | / |- folded/literal block scalars). Tracked for v0.29.1 patch.
• Production build OOMs on this 3.8 GB box during the PWA service-worker step. Workaround used for v0.29.0: stop the old prod container (~50 MB), run BUILD_TARGET=prod docker compose build web, restart. 2.26 GB available post-kill was enough. If we add more content the threshold may break again — long-term fix is to build on GitLab CI (no memory limit) or add swap.

📚 Files changed (12)

• package.json + pnpm-lock.yaml — mermaid added
• src/app.css — mermaid + primer + glance styling
• src/routes/tech/+page.server.ts — marked renderer override for mermaid passthrough
• src/routes/tech/+page.svelte — mermaid client-side initialise + run on mount
• src/routes/tech/tech.test.ts — +47 tests for US41
• src/content/tech/01-satellite-hardware.md — restructured
• src/content/tech/02-satellite-communication.md — restructured
• src/content/tech/03-ai-hardware.md — restructured
• src/content/tech/04-ai-software.md — restructured
• src/content/tech/05-scaling-laws.md — restructured
• src/content/tech/06-jensens-5-layer-cake.md — restructured
• src/content/tech/07-karpathys-software-eras.md — restructured
• static/tle/kuiper.json + static/tle/oneweb.json — auto-refreshed by daily prebuild cron

✅ Verified

• pnpm test:unit — 394/394 green (was 347 before US41)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean
• pnpm format — clean
• BUILD_TARGET=prod docker compose build web — succeeded after killing the old prod container to free RAM (OOM workaround, see Known Issues)
• Live container high-machines-web — running v0.29.0, healthy, /tech renders Primer + Mermaid + Examples + At-a-glance + In-depth for all 7 sections
• PWA still mandatory (per JP 2026-06-17 10:50 UTC) — full workbox precache shipped (130 entries, ~6 MB)

[0.28.1] - 2026-06-17

🔭 Vision

Double-check of US40 (v0.28.0) found one factual error and propagated it across two files. Patch-bumped (no new feature). The fix: the LEO orbit period + eclipse-fraction claim in §6 and §1 was wrong (it conflated the orbital period with the sunlit fraction), and the fix adds primary-source citations to back the corrected numbers. All other US40 claims (Huang's 5 layers, Huang's quotes, Karpathy's 3 eras, Karpathy's quotes) were re-verified against primary sources and are accurate.

🪛 Fixed

• LEO orbit-period claim in 06-jensens-5-layer-cake.md §6 and 01-satellite-hardware.md §1. Original: "≈55-min orbits with ≈30 min of eclipse". That conflated the ~55 min of sunlight per orbit with the ~90 min orbital period, and under-stated the eclipse fraction. Corrected to: "~90-min orbital period with ~35 min of eclipse and ~55 min of sunlight per orbit". Added primary-source citations to the ScienceDirect LEO overview ("approximately 65 min in the Sun and eclipsed for 35 min") and the ERAU IJAA paper on LEO eclipse computation ("in LEO the maximum eclipse duration remains close to 35 minutes"). Same correction applied to both files; the error was pre-existing in §1 (US32 / v0.22.0) and was propagated into §6 (US40 / v0.28.0) by my US40 cross-reference paragraph.

📚 Verification report (US40, double-check 2026-06-17)

Per JP's request after US40 ship ("Please double check your work"), re-verified every factual claim in the two new sections against primary sources:

• §6 Jensen Huang's 5-layer cake — all direct quotes verified verbatim against the NVIDIA blog post by Jensen Huang, 10 March 2026 (confirmed byline in <meta name="author" content="Jensen Huang" /> and the page's datePublished: 2026-03-10T10:00:03+00:00 JSON-LD). Layer order energy → chips → infrastructure → models → applications matches the original verbatim. All five in-text quotes (energy / chips / infrastructure / models / applications, plus the "every successful application pulls on every layer..." closer) match the source character-for-character. The "a few hundred billion dollars into [the buildout]; trillions of dollars" quote is accurate (paraphrased from "We are a few hundred billion dollars into it. Trillions of dollars of infrastructure still need to be built."). Secondary source ASA Computers reproduces the same layer order.
• §7 Karpathy's software eras — all direct quotes verified against the 2017 Software 2.0 essay (date 11 November 2017 confirmed in the page's metadata: "9 min readNov 11, 2017") and the June 2025 YC AI Startup School talk transcript on singjupost (date 18 June 2025 confirmed in the byline). All five in-text quotes match the sources character-for-character. The Latent Space annotation ("Software 3.0 is eating 1.0/2.0 — a huge amount of software will be rewritten") and MindStudio FAQ ("one where the program is compiled from data rather than written by hand") are verified verbatim against the Latent Space annotation and MindStudio summary, 2 May 2026 (date confirmed via the page's datePublished: 2026-05-02T00:00:00.000Z JSON-LD). The YC library page exists with the correct title.
• All cross-reference targets (§1, §2, §3, §4, §6, §7) verified to exist by the pnpm test:unit slug test (347/347 green).
• Math checks — 10 kW / (30% BOL efficiency × AM0 1361 W/m²) = ~25 m² of triple-junction GaAs solar arrays. Confirmed against CAVU Aerospace 30% triple-junction GaAs cell and 32% BOL datasheets. Within engineering uncertainty (cells degrade ~15% over 5 years; 25 m² is the BOL figure for 10 kW continuous).
• The one error found was the LEO orbit-period conflation (above).

⚠️ Known Issues

• The §1 satellite-hardware LEO battery-bank line ("battery bank capable of delivering 5 kW for 35 minutes") was kept as-is in this patch. The corrected math is that a 10 kW continuous payload needs ~5.8 kWh (10 kW × ~35 min eclipse), not the 2.9 kWh implied by 5 kW × 35 min. The §1 wording could be read as "5 kW battery for a 10 kW payload" (under-spec for continuous) or as "5 kW average payload in eclipse mode" (intentional throttle) — JP to clarify in a follow-up. Out of scope for v0.28.1.

[0.28.0] - 2026-06-16

🔭 Vision

The /tech page now has the two AI frameworks every technical investor expects: Jensen Huang's "5-layer cake" (NVIDIA's AI factory: Energy → Chips → Infrastructure → Models → Applications) and Andrej Karpathy's "software 1.0 / 2.0 / 3.0" (human code → neural network weights → LLM prompts in natural language). Both add depth to the existing 5 sections without changing the page's structure. The two sections cross-reference the existing AI hardware + AI software sections, so a reader sees how Jensen Huang's stack frames the same hardware, and how Karpathy's eras classify the same models by how they are programmed.

✨ Added

• New /tech#06-jensens-5-layer-cake — Jensen Huang's 5-layer cake (NVIDIA blog, 10 March 2026: https://blogs.nvidia.com/blog/ai-5-layer-cake/). The 5 layers in his exact order: energy → chips → infrastructure → models → applications. The orbital case is concrete: Huang identifies energy as the only binding constraint on orbital AI, and applications as the only layer where economic value is created. The section ties each layer to the existing 5 sections (e.g. energy → §1 satellite hardware + solar arrays; chips → §3 AI hardware; infrastructure → §2 satellite communication + Walker constellations; models → §4 AI software; applications → real-time inference close to the user).
• New /tech#07-karpathys-software-eras — Andrej Karpathy's three eras. Primary sources: the 2017 Software 2.0 essay (Software 1.0 = Python/C++ written by humans, Software 2.0 = neural network weights compiled from data) and the June 2025 YC AI Startup School talk (Software 3.0 = LLM prompts in natural language — "remarkably we're now programming computers in English"). Cross-references the same §3 + §4 sections, framing the orbital implication: a Software 3.0 stack on orbit is upgradable without an upload — the prompt is the program, not the weights.
• Both new sections carry inline markdown citations [label](url) to primary + secondary sources (NVIDIA blog, Karpathy's 2017 essay, YC AI Startup School transcript, Latent Space annotations, MindStudio summary, secondary tech-press coverage). Every key claim is cited from at least one primary source.

🔧 Changed

• src/routes/tech/+page.svelte intro cell — table now has 7 rows (was 5); intro paragraph updated to "seven technology layers"; the comment in the svelte file is now "+ 7-row static table".
• src/content/tech/03-ai-hardware.md — added "See also" cross-reference to §6 (the chip layer is Huang's layer 2 of 5; orbital DC chip choices are gated by the energy layer above).
• src/content/tech/04-ai-software.md — added "See also" cross-references to both §6 and §7 (the inference engines + model architectures are Huang's layer 4 of 5; the same models are programmed three different ways per Karpathy).
• src/routes/tech/tech.test.ts — expected slug list extended from 5 to 7; id={section.slug} checks cover the 2 new sections.

🪛 Fixed

None this release.

⚠️ Known Issues

• The 2 new sections don't have "Primer" + mermaid diagrams + examples/emojis yet. That enrichment is US41 (in planning 2026-06-16). For now the new sections are plain markdown like the original 5.
• The emojimap (🛰️📡⚡🧠📈🏗️🔮) is not yet applied. US41 will add leading emojis to each section.

✅ Verification

• 347/347 unit tests green (unchanged from v0.27.1 — the existing tests now cover 7 slugs instead of 5; no new tests added but the existing tests would have failed if the new files weren't in place)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean
• pnpm build — succeeds
• Container rebuilt with BUILD_TARGET=prod, healthy on https://localhost:8001/tech#06-jensens-5-layer-cake and ...#07-karpathys-software-eras
• Commit: e922ecb (US38) + 7b27219 (v0.27.0 closeout) + f79cacc (US39, v0.27.1) + ff3bdce + a309f2b (US40 + US41 planning) + (this commit, v0.28.0)

[0.27.1] - 2026-06-16

🔭 Vision

The middle of BACKLOG.md — the per-US detail blocks — is now as scannable on mobile as the top of the file (US38). US39 restructured the 5 worst wall-of-text lines (single-paragraph blobs of 500–714 chars) into bullet points with sub-bullets for file paths, code snippets, and command examples. The other ~10 long lines in the middle (Sprint 2/3/4 closed summaries, US31 detail block, US14 Sprint 2 review notes) stay as-is for now — they follow different restructure patterns (intentional summaries, sub-section headings + dense bodies) and are tracked as US40+.

✨ Added

None this release.

🔧 Changed

• BACKLOG.md US2 Implementation notes — "Styling via Tailwind" (520 chars) → 1 parent bullet + 2 sub-bullets (Workaround + Follow-up)
• BACKLOG.md US10 Implementation notes — "30-minute rabbit hole" (662 chars) → 1 parent bullet + 3 sub-bullets (Result + The fix + Finding it)
• BACKLOG.md US15 post-mortem — "ECharts init" (662 chars) → 5 bullet points
• BACKLOG.md US35 Why this matters — "missing link between config" (568 chars) → 4 bullet points
• BACKLOG.md US36 T36.21 — "Visual smoke done" (714 chars) → 1 parent checkbox + 4 sub-bullets
• BACKLOG.md Quick Stats — 32 US shipped (was 31), 347 unit tests (unchanged)

🪛 Fixed

None this release.

⚠️ Known Issues

• ~10 more long-paragraph lines remain in the middle of BACKLOG.md — Sprint 2/3/4 closed summary blocks (describing the US1-US27 ships), US14 Sprint 2 review notes (US25/US26 ship summaries), US31 detail block (Phase 4 + Secondary wins), US30 + US34 minor long lines. Each follows a different restructure pattern. Tracked for US40+.

✅ Verification

• 347/347 unit tests green (unchanged from v0.27.0 — no code change, sanity check)
• 6/6 e2e tests green (unchanged from v0.27.0)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean
• pnpm build — succeeds
• Container rebuilt with BUILD_TARGET=prod, healthy on https://localhost:8001, version footer reads v0.27.1
• Commits: e922ecb (US38, shipped as v0.27.0) + 7b27219 (v0.27.0 closeout) + (this commit, v0.27.1)

[0.27.0] - 2026-06-16

🔭 Vision

The /info page's BACKLOG tab is now scannable on mobile. JP's complaint ("way too dense for easy readability") was structural, not stylistic: 7 lines in the top of BACKLOG.md were over 1,000 chars, with the "Last US36 ship" line at 2,516 chars of **bold:** text separated by em-dashes. The /info page now renders the top of the BACKLOG as a short headline + a per-US "Recent ships" table-of-contents (one H4 per US, each with a 2-3 line summary + "Files changed" + "Verified" bullet lists), so JP can read "what shipped, what files changed, what was verified" at a glance on a phone.

✨ Added

• New ### Recent ships H3 in BACKLOG.md with one #### USxx H4 per US (US37, US36, US35, US34, US33, US31, bug fixes) — each with a 2-3 line summary + **Files changed (N):** + **Verified:** bullet lists
• New ### Recent planning H3 for the older planning entries (US35 update, US36 update, US35 scope expansion)
• New scripts/us38-restructure-top.py — Python script that does the literal-string replacement of the old dense block with the new structured block (used because the edit tool's exact-text-match was fragile for a 10,977-char dense block)

🔧 Changed

• BACKLOG.md — top "Project Status Overview" block restructured: 13-line dense block (with lines of 800–2,516 chars) → 100-line scannable structure (no line > 500 chars in the replaced section)
• BACKLOG.md — Quick Stats updated (31 US shipped, 37 US total, 2 in planning)
• package.json — version bumped 0.26.0 → 0.27.0
• SPEC.md line 3 — status line updated to v0.27.0 (was stale at v0.26.0)
• README.md line 5 — status line updated to v0.27.0 (was stale at v0.24.0 — a pre-existing bug, not caused by US38)

🪛 Fixed

• README.md line 5 status line was stale at v0.24.0 — never bumped for v0.25.0 (US28) or v0.26.0 (US37). Now correctly reads v0.27.0.

⚠️ Known Issues

• The middle of BACKLOG.md (per-US H2 detail blocks like ## US1:, ## US2:) still has the single-paragraph-blob problem. This US fixed the top of the file (the worst pain), not the middle. A future US (US39+) should restructure the per-US detail blocks too.

✅ Verification

• 347/347 unit tests green (unchanged from v0.26.0 — no code change, sanity check)
• 6/6 e2e tests green (unchanged from v0.26.0)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean
• pnpm build — succeeds
• Container rebuilt with BUILD_TARGET=prod, healthy on https://localhost:8001, version footer reads v0.27.0
• Live /info#backlog: confirmed Recent ships and Recent planning H3s present in the rendered HTML
• Commits: e922ecb (US38 implementation) + (this commit, v0.27.0 closeout)

[0.26.0] - 2026-06-16

🔭 Vision

A dedicated /info page that renders the 4 project markdown files (Backlog, README, Spec, Changelog) in a single focused view with deep-linkable URL-hash tabs. JP can now read the full project state — what's planned, what's shipped, what's in the design doc, what's in the release notes — without leaving the site or juggling 4 editor tabs. Default tab is changelog (most relevant for a returning visitor).

✨ Added

• New /info route with 4 tabs (Backlog, README, Spec, Changelog)
• New src/routes/info/+page.server.ts reads all 4 markdown files at SSR time via readFile + marked.parse()
• New src/routes/info/+page.svelte renders the active tab's markdown with {@html} over a Tailwind prose prose-invert body
• URL-hash tab state (#backlog, #readme, #spec, #changelog) for deep linking — read on mount + listened via hashchange event
• Anchor links for every ## H2 heading (slugified id attributes via a 5-line addHeadingIds helper — no new dep)
• Per-page <title> ("Project state — high machines") + <meta description>
• 9 new tests in src/routes/info/info.test.ts (route wiring, server load reads 4 files, 4 tabs in template, URL hash state, default tab = changelog, structural anchors, title + meta, {@html} rendering, heading-IDs helper)

🔧 Changed

• SPEC.md §3 — /info moved from "v1+ planned" to "v1 routes" (the route ships as part of v0.26.0)
• SPEC.md revision bumped from v0.12 → v0.13 (new route added to the IA)
• BACKLOG.md — US37 status updated to ✅ Done; Quick Stats bumped (30 US shipped, 36 US total, 2 in planning)
• package.json version bumped 0.25.0 → 0.26.0
• BACKLOG.md "Last Updated" + "Last US ship" lines reflect US37 ship

🪛 Fixed

None this release.

⚠️ Known Issues

• The {@html} rendering is unsanitized — the markdown is project-controlled (no user input), but a dompurify pass is a documented follow-up for defense in depth (per US37 implementation notes). The svelte/no-at-html-tags ESLint rule is disabled in the script block with a comment explaining the trust boundary.
• The /info page is public — anyone with the URL can read the BACKLOG, README, SPEC, and CHANGELOG. Gating (auth, IP allowlist, or hidden from Nav) is a follow-up US if JP wants it private.
• TLE cache files are stale (carried from v0.25.0) — static/tle/kuiper.json and static/tle/oneweb.json are unmodified from a prior TLE fetch (Starlink is regularly 403-rate-limited by CelesTrak). The prebuild hook runs pnpm tle:fetch best-effort (|| echo) so a rate-limit doesn't break the build — the consequence is stale cache.

✅ Verification

• 346/346 unit tests green (was 337 at v0.25.0, +9: all in info.test.ts)
• 6/6 e2e tests green (unchanged from v0.25.0)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean (Prettier + ESLint flat config)
• pnpm build — succeeds
• Container rebuilt with BUILD_TARGET=prod, healthy on https://localhost:8001, version footer reads v0.26.0
• Dev-server smoke test: /info renders with 4 tabs, default tab = changelog, URL hash switches active tab, all 4 markdown files render with anchor IDs on <h2> headings
• Commits: 6545bff (US37 planning) + (this commit, US37 implementation)

[0.25.0] - 2026-06-16

🔭 Vision

First dedicated detail pages for each of the 11 tracked competitors. Visitors and investors can now click into a curated digest of any competitor (name, oneLiner, summary, highlights, watchlist, sources) instead of hitting a 404. The website's research depth is now browsable, not just listed — each card on /competitors links to a real page, and each page has its own SEO surface (<title> + <meta description>) so social shares stop showing a generic SvelteKit fallback.

✨ Added

• New /competitors/[slug] detail route — name, oneLiner, summary, highlights, watchlist, sources, "Back to competitors" link, "Source on GitLab" raw-markdown link
• New typed data module src/lib/competitors/data.ts — CompetitorProfile interface + 11 curated digests + getCompetitorProfile() + formatReviewedDate() + listPublishedCompetitors() + rawMarkdownUrl() helpers
• New "Reviewed YYYY-MM-DD" date on every published list card
• "Source on GitLab" link on every detail page (transparency affordance for the curious skeptic)

🔧 Changed

• /competitors list page rewired to read from the typed data module (no more auto-render from raw research markdown)
• Detail page sets per-profile <title> and <meta description> (pre-empts half of US20 SEO work)
• BentoCell layout used for the detail page (header, summary, highlights, watchlist, sources) — consistent with /competitors, /tech, /explore

🪛 Fixed

• "View profile →" link on each list card was a 404 / pointed to the raw markdown — now points to a real detail page
• Mantis-space direct URL access (e.g. /competitors/mantis-space) now 404s, consistent with the list filter (no "coming soon" placeholders)

⚠️ Known Issues

• mantis-space digest is still pending review (last research update 2026-03-16) — hidden from the list + 404 on direct URL. Change verification: "Pending review (last research update 2026-03-16)" → verification: "Verified YYYY-MM-DD" in data.ts to publish it.
• TLE cache files are stale — static/tle/kuiper.json and static/tle/oneweb.json are unmodified from a prior TLE fetch (Starlink is regularly 403-rate-limited by CelesTrak). The prebuild hook runs pnpm tle:fetch best-effort (|| echo) so a rate-limit doesn't break the build — the consequence is stale cache.

✅ Verification

• 337/337 unit tests green (was 318 at v0.24.0, +19: 11 in data.test.ts + 8 in [slug].test.ts)
• 6/6 e2e tests green (unchanged from v0.24.0)
• pnpm check — 0 errors / 0 warnings
• pnpm lint — clean (Prettier + ESLint flat config)
• pnpm build — succeeds
• Dev-server smoke test: list 10 cards (mantis-space hidden), detail /competitors/spacex renders all 6 structural anchors, 404 on /competitors/mantis-space + /competitors/nonexistent
• Container rebuilt with BUILD_TARGET=prod, healthy on https://localhost:8001, version footer reads v0.25.0
• Commits: 13d2b18 (feat: US28) + fcc6366 (fix: 404 for pending profiles)

📚 Legacy entries (v0.1 – v0.24)

Pre-restructure format. The original prose entries are preserved verbatim for historical record — these are the in-the-trenches writeups from the day each version shipped, with the full commit chain, test counts, and verification evidence. The new grouped format (Vision / Added / Changed / Fixed / Known Issues / Verification) is applied to v0.25.0 above and will be used for everything going forward.

• v0.24 (2026-06-15): US35 + US36 shipped. Orbit visualization — the original ringsData-renders-a-halo bug is fixed. The actual fix was a unit-mismatch (km → Earth-radii) at 3 call sites (c4830a0: 982 LEO sats at alt=550 km were rendered at 550× Earth-radii = a flat plane off in space, perspective-projected as a fan of straight lines from the camera; new kmToGlobeUnits helper in globe-data.ts applied at 3 call sites; 2 new failing tests in globe.test.ts). A second fix (3dca294): explicit pointOfView on the Globe instance (lat: 15, lng: -10, altitude: 2.8 Earth-radii) for consistent desktop/mobile framing, plus thicker orbit stroke (0.5 → 1.0) and finer path resolution (default 2°/seg → 1°/seg). US35 commits shipped the math (computeReferenceOrbitPath), the per-file config architecture (6 satellite configs in src/lib/satellites/config/), the GeoJSON ground-stations loader, and the makeMarker symbol layer. US36 commits: 76cb4fd (mesh factory wired into both globes via customLayerData + customThreeObject; pointLabel/pointColor/pointRadius retired), 409607f (e2e fixtures — axiom-1 orbital DC + reunion-island ground station; per-file config + GeoJSON patterns verified end-to-end with one new file each, no other code touched), e612ae0 (8 new unit tests for the orbital-DC marker shape — THREE.Group with sphere + LineSegments panel), f410c01 (per-marker userData.testid + polar camera cap controls.minPolarAngle = 0.25 + maxPolarAngle = π - 0.25 + window.__globe debug handle for e2e walks), c64e453 (catalog toggle list bug fix — toggle list now always rendered, count + Loading state stays inside the conditional), 2ca9ca9 (Playwright e2e suite — e2e/globe-orbits.spec.ts + e2e/per-file-config.spec.ts, 6/6 green). The e2e suite walks the three.js scene via window.__globe to assert: 8 reference orbits in pathsData (regression gate for the original bug), no orphan halos, and 4 fixtures present on the live globe (axiom-1, starcloud-1, ksat-svalbard, reunion-island). 318/318 unit tests green (was 303 at v0.23.0, +15: 8 orbital-dc-marker + 1 catalog toggle + 1 explore-globe regression + 5 constellation-config). 6/6 e2e tests green (new). pnpm check 0/0, pnpm lint clean, pnpm build clean, full prod rebuild + boot + visual smoke. Live site shows 983 sats (982 from T36.7 + 1 axiom-1 fixture) + 5 ground-station networks + orbital-DC meshes (Starcloud-1, axiom-1) at their propagated positions + 8 closed elliptical reference orbits. The version footer at the bottom-right of every page now reads v0.24.0.
• v0.23 (2026-06-11): US34 shipped + 2 bug fixes. US34 container build profiling — pnpm profile:build wraps docker compose build --progress=json, parses per-stage BuildKit vertex durations (BuildKit 0.11+ / docker 23+ format), writes data/build-profile.json (gitignored, 10-rotated via PROFILE_KEEP), prints a human-readable summary + a diff vs the previous run. 16 new vitest tests in scripts/profile-build.test.ts (parser noise tolerance, malformed-JSON tolerance, stage grouping, cached-step counting, summary rendering, diff flagging of added/removed stages and +N cached deltas). Env vars: BUILD_TARGET (dev/prod), NO_CACHE=1 (cold build), PROFILE_KEEP=20 (more history). No Dockerfile change, no new dependencies, no third-party tool. The diagnostic playbook (5 commands, works today without US34) and the standard mitigations list are documented in README §Build profiling, SPEC §7, and BACKLOG US34. Verified live: dev rebuild 1.6 s, prod cold build 173.2 s (builder 105.1 s + prod 6.8 s), image 155 MB / 9 layers. Bug fixes 2026-06-10 (also part of v0.23): Globe $state(true) reactivity fix (17c0c26); Vite 6 server.allowedHosts fix for Traefik host-header 403s (85c7d1f). 272/272 unit tests green at v0.23.0 close.
• v0.22 (2026-06-11): US32 shipped — /tech explainers page (5 sections, markdown-driven, bento grid, deep-linkable anchors). New src/content/tech/ directory with 01-satellite-hardware.md, 02-satellite-communication.md, 03-ai-hardware.md, 04-ai-software.md, 05-scaling-laws.md. Each has frontmatter (title, subtitle, whyForOrbitalDC, keyTerms: [..]) + a 600-800 word prose body. New src/routes/tech/+page.server.ts parses the frontmatter + renders the body with marked (already a US5 dep). New src/routes/tech/+page.svelte renders an intro bento cell (1-paragraph summary + 5-row static table of the 5 layers) + 5 section cells, each with prose + an amber-coloured "Why for orbital DC" callout + a key-terms sidebar. BentoCell.svelte gained an id prop for the section anchor (used as /tech#<section> deep links). Cross-references back: /story/02-problem.md → /tech#01-satellite-hardware + /tech#03-ai-hardware; /explore footer → /tech#02-satellite-communication + /tech#01-satellite-hardware; /competitors footer → /tech#03-ai-hardware + /tech#02-satellite-communication. The Nav already had the "Tech" entry (it was added during US9 in anticipation). 10 new tests in src/routes/tech/tech.test.ts covering: route wiring, content directory + 5 expected files, frontmatter shape, page wrapper testid, section-cell id matching the slug (bound or literal), intro static table, nav entry, cross-refs from /story + /explore + /competitors. pnpm test:unit 256/256 green (was 246, +10). pnpm check 0 err / 0 warn (the pre-existing webglOk warning cleared with the cleanup). pnpm lint clean. pnpm build succeeds. Out of scope (flagged for future work): the lighthouse-baseline runner is still pointed at the original 4 routes; adding /tech to the runner is a small follow-up tracked in the BACKLOG T32.9 task. Out of scope (deferred): verification-pass on each section's cited claims — the existing 5 sections are paraphrases of the primary sources already cited in 01-research/, but the dedicated verification sweep (one sub-agent per section, applying the corrections delta) is a follow-up before the page is shown to investors.
• v0.21 (2026-06-09): US18 shipped — Sprint 4 closed. New src/lib/observability/ module with a tiny wrapper around @sentry/svelte (@sentry/svelte@10.57.0 added as a dependency). The wrapper is env-gated: every public function (init, captureError, capturePageView, getAnonymousId) is a no-op when PUBLIC_SENTRY_DSN is empty. The SDK is dynamically imported inside init() so Vite/Rollup can tree-shake the whole SDK away when the DSN is unset — bundle size is unchanged when the DSN is empty. SSR-safe: every public function guards on typeof window !== "undefined" so Node-side rendering never loads the browser-only SDK. Anonymous user ID: a random UUID v4 stored in localStorage under high-machines-anonymous-id (the old generic key was rebranded along with everything else). Wired into src/routes/+layout.svelte: init() called on mount, capturePageView(page.url.pathname) runs as a $effect on every route change, and global window.onerror + window.onunhandledrejection handlers capture any error that escapes. 12 new tests in src/lib/observability/observability.test.ts covering: scaffolding (file exists, exports are present, SSR guards, env-gating, dynamic import, localStorage key); unit tests (init is a no-op, captureError is a no-op, capturePageView is a no-op, getAnonymousId returns a UUID v4); layout wiring; .env.example documentation. pnpm test:unit 216/216 green (was 204, +12). pnpm check 0 err / 1 pre-existing webglOk warning. pnpm lint clean. pnpm build succeeds. Manual smoke (pending JP): a real browser session with a real DSN to confirm the breadcrumb trail and error capture work end-to-end. What this is NOT: a full analytics suite, a session-replay tool, a push-notification system. Sprint 4 closed: US15 + US16 + US17 + US18 + US26 + US27 all done.
• v0.20 (2026-06-09): US27 shipped — rebrand to "high machines" with the slogan "Impossible. But doable." The public-facing brand string is now consistent across: the PWA manifest (name, short_name, description in vite.config.ts); all page <title> tags; the landing-page hero (<h1>high machines</h1> + the slogan in a small monospace sky-300 label below the h1); the Nav brand text in the sticky top bar; the offline page title + message; the PWA install banner message; and all aria-labels. The PWA install name is now high machines (so the app-icon label on the home screen will say "high machines"). New tests: src/lib/components/nav/Nav.test.ts gained a US27 assertion; src/lib/pwa.test.ts asserts "high machines" in the manifest and the absence of "Space-AI Website"; src/routes/landing.test.ts asserts the h1, the slogan, and the new <title>. 4 new test assertions (1 in Nav, 1 in pwa, 3 in landing, 1 in nav) — pnpm test:unit 206/206 green (was 200, +6). pnpm check 0 err / 1 pre-existing webglOk warning. pnpm lint clean. pnpm build succeeds. Out of scope (flagged for JP): the folder name 01-space-AI/07-website/ is unchanged — that is a monorepo structural decision with a bigger blast radius (docker image name, container name, docker-compose.yml service references). The brand string and the code-org string are now intentionally different; this is a 0.25-day follow-up if JP wants to align them. Sprint 4 now 5/5 + US26 ✅.
• v0.19 (2026-06-09): US16 shipped — PWA install UX (install banner, update toast, offline fallback page). New src/lib/pwa/ module: PwaInstallBanner.svelte (listens to beforeinstallprompt, shows a non-modal banner with Install + Not now buttons, persists dismissal in localStorage under pwa-install-dismissed); PwaUpdateToast.svelte (uses useRegisterSW from virtual:pwa-register/svelte, shows a bottom-right toast with Reload + Later buttons when a new SW version is detected). New /offline route (src/routes/offline/+page.svelte) — Workbox's navigateFallback: "/offline" serves this page when a navigation request fails AND the response isn't in the precache AND the network is down. vite-plugin-pwa switched from registerType: "autoUpdate" → registerType: "prompt" (so the toast can intercept the update). navigateFallbackDenylist: [/^\/api\//, /^\/sw\.js$/] keeps API endpoints and the SW itself out of the fallback. Both PWA components mounted in src/routes/+layout.svelte (install banner above main content, update toast below the version footer). workbox-window added as a devDep (v7.4.1, MIT) — required by the prompt-mode useRegisterSW path. New src/lib/pwa/virtual-pwa.d.ts type declaration file (the virtual module is build-time only, so svelte-check needs a hand-rolled declaration). 18 new tests across 4 test files (5 in pwa.test.ts — wiring, manifest, icon, prompt mode, navigateFallback; 5 in pwa-update-toast.test.ts; 7 in pwa-install-banner.test.ts; 4 in offline.test.ts). pnpm test:unit 200/200 green (was 182, +18). pnpm check 0 err / 1 pre-existing webglOk warning. pnpm lint clean. pnpm build succeeds. Bugfix: the onRegisteredSW callback signature in PwaUpdateToast.svelte was failing pnpm check with a type error — the actual vite-plugin-pwa types pass ServiceWorkerRegistration | undefined, not ServiceWorkerRegistration. Fixed by widening the parameter type. Sprint 4 now 4/5 + US26 ✅.
• v0.18 (2026-06-09): US15 + US17 shipped — (US15) Lighthouse pass: all 4 routes ≥ 94 on best-practices + a11y + perf. The ECharts wrapper (<Chart />) was replaced by a zero-JS <StaticBarChart /> (pure HTML/CSS horizontal-bar chart with auto log/linear scale). The /explore live chart is now a $derived reactive StaticBarChart (only the current-altitude bar's colour changes — Tailwind class binding, not a full chart re-render). ECharts npm dep removed from package.json (no consumers). a11y sweep: text-slate-500 → text-slate-400 (color contrast), <h3> → <h2> in bento card titles (heading order), min-h-11 min-w-[7rem] on the GS-cone + constellation chips (touch-target), underline on the "US13" link (link-in-color-block). New pnpm lighthouse:baseline script (scripts/lighthouse-baseline.mjs + lighthouse + chrome-launcher devDeps, reuses the Playwright Chromium that the test runner already pins) exits 1 when any route scores < 90 in any of the 3 categories. Refresh-run scores (Lighthouse v13.3.0, mobile profile, prod preview, after the prod container was rebuilt): / 98/100/100, /story 94/100/100, /competitors 97/100/100, /explore 96/100/100. pnpm test:unit 182/182 green (was 194, –12 from removing the ECharts wrapper tests; the new StaticBarChart + live-chart tests replaced them). pnpm check 0 err / 1 pre-existing webglOk warning. pnpm lint clean. pnpm build succeeds. ECharts can be re-added from git history if a future chart needs interaction that the static component can't express. (US17) Production-Docker + JP's-VPN deployment hardening: the high-machines-website:prod container is the actual runtime (verified via docker ps, healthy on high-machines-web); Traefik fronts it for HTTPS (US4); JP's VPN is the access path. Cloudflare quick-tunnel idea was dropped per JP (not European); all cloudflared references removed from README.md + BACKLOG.md. The "Docker-only. HTTPS-only." hard-constraint callout is unchanged. Sprint 4 now 3/5 + US26 ✅.
• v0.17 (2026-06-08): US26 shipped — curated ground-station networks with visibility cones + GB/day compute on /explore. New src/lib/ground-stations/ module: typed catalog (GroundStationNetwork + GroundStationEntry interfaces) with 8 European-leaning networks (KSAT, Atlas, Leaf Space, Viasat RTLS, Goonhilly, Azercosmos, Hisdesat, OHB) × 23 sites. Per-site great-circle visibility cones (radius = arccos(R_earth / (R_earth + altitudeKm)) in radians, rendered as small circles on the globe). New <ExploreGlobe> overlay section with: network filter (checkbox per network, color-dot matches markers, Hisdesat defaults off as government/military), per-site row with cone toggle chip, reset button, localStorage persistence under groundStations.enabledNetworks. 2 new data-rate sliders (Downlink 0.1–10 Gbps, Uplink 0.01–1 Gbps) wired into params.svelte.ts. New GB/day compute widget: orbital-regime visibility fraction (LEO 0.10, MEO 0.30, GEO 0.95) × per-sat Gbps × network scale factor (saturating at 0.5). coneRadiusFromAltitude and computeGBPerDay are pure functions for TDD. 30 new tests (catalog shape, cone math at 400/20000/35786 km, GB/day compute, globe integration, params wiring); pnpm test:unit 194/194 green (was 164, +30). pnpm check 0 err / 1 pre-existing webglOk warning. pnpm lint clean. pnpm build succeeds. Sources for every site documented in 01-space-AI/01-research/ground-station-networks.md. Required fix: vite.config.ts adds optimizeDeps.esbuildOptions.alias for #wasm-single-thread / #wasm-multi-thread (in addition to the existing resolve.alias and ssr.external) — without it, the dev server's esbuild pre-bundler tried to bundle the satellite.js v7 pthreads IIFE which uses top-level await. pnpm build was always fine; only vite dev was hitting it.
• v0.16 (2026-06-08): US25 shipped — local satellite catalog (OMM-first, on-demand API). The 6 default constellations (ISS, Starcloud-1, CSS TIANHE inline-OMM; OneWeb, Kuiper, Axiom celestrak-prebuild) are now committed to src/lib/satellites/catalog.ts and drive both the build-time cache script (scripts/fetch-tle.ts is now catalog-driven: filters source === "celestrak-prebuild" and skips inline-OMM entries) and the runtime enable/disable UI in <ExploreGlobe>. Inline-OMM entries (single satellites) ship in the catalog and are refreshed by the new pnpm tle:update-inline script (scripts/update-inline-omm.ts); prebuild entries still come from CelesTrak via pnpm tle:fetch. New /api/celestrak/[group] SvelteKit endpoint (src/routes/api/celestrak/[group]/+server.ts) proxies CelesTrak for runtime-enabled constellations (whitelist: Starlink, OneWeb, Kuiper, stations, GNSS, GPS, Galileo, GLONASS, Iridium, Planet, Axiom, etc.) with a 1h in-memory cache and a 15s fetch timeout. The ExploreGlobe overlay now has a data-testid="constellation-toggle-list" button per catalog entry — clicking a celestrak-runtime entry calls the runtime API and adds the sats live; enabled set persists in localStorage under satell…dIds. New $lib/satellites/runtime.ts helper for the client-side fetch. The US14 detail block in BACKLOG.md is annotated "Superseded by US25". 12 new tests (catalog shape + script shape + API shape + static cache validation); pnpm test:unit 164/164 green. pnpm check 0 err / 0 warn. pnpm lint clean.
• v0.15 (2026-06-07): US14 shipped — TLE/OMM-based satellite tracking. The build-time cache script (scripts/fetch-tle.ts) fetches BOTH OMM JSON (FORMAT=json, primary path via json2satrec) and legacy 3LE text (FORMAT=tle, fallback via twoline2satrec) from CelesTrak, merges per-NORAD_CAT_ID, and writes a single static/tle/<group>.json per constellation. propagatePosition auto-detect: OMM wins if present, TLE is the fallback. Wired into ExploreGlobe.svelte — real constellation markers render on the globe; per-constellation count + nearest-orbit badge in the overlay. pnpm tle:fetch is a standalone script; prebuild runs it best-effort (|| echo) so a CelesTrak rate-limit doesn't break the build. tsx added as a devDependency. 14 new tests in tle.test.ts + 7 in fetch-tle.test.ts; pnpm test:unit 147/147 green. Note: Starlink is regularly 403-rate-limited by CelesTrak; the 4 other groups (OneWeb, ISS, Kuiper, GNSS) consistently succeed.
• v0.14 (2026-06-08): US13 shipped — 3D orbit viz on /explore. <ExploreGlobe> wraps Globe.svelte with reactive ringsData from exploreParams.safeParams.altitudeKm via $derived.by; amber ring for the nearest reference orbit, sky for the other 7. Globe.svelte gains a ringsData prop (backward compat: falls back to leoOrbitRings when omitted). Constellation overlay badge shows sat count, GS count, nearest orbit name. 5 new TDD tests; pnpm test:unit 125/125 green (was 120, +5). pnpm check 0/0. pnpm lint clean. Live on https://localhost:8001/explore (globe cell, amber ring on LEO). Sprint 3 in progress (4/5).
• satellite.js v7.0.1 upgrade (2026-06-07): Upgraded from v6.0.2 → v7.0.1. v7 ships WASM runtimes (single/multi-thread) that use top-level await in an IIFE context — Rollup's IIFE output format doesn't support that. Workaround: vite.config.ts adds resolve.alias pointing #wasm-single-thread and #wasm-multi-thread to src/lib/stubs/empty-wasm.js (a 4-line stub). The website only uses the pure-JS SGP4 API (twoline2satrec, propagate, gstime, json2satrec, etc.) so the stub has zero runtime effect. TypeScript check clean, build succeeds.
• v0.13 (2026-06-07): US11 shipped — live latency chart on /explore. <ExploreLatencyChart> re-derives its ECharts option from exploreParams.safeParams.altitudeKm via $derived; the US8 Chart wrapper's $effect handles the live setOption. The bar nearest the current altitude is highlighted amber, the others sky-400. 5 new TDD tests; pnpm test:unit 120/120 green (was 115, +5). pnpm check 0/0. pnpm lint clean. Live on https://localhost:8001/explore (56 KB, chart cell visible). Sprint 3 in progress (3/5).
• v0.12 (2026-06-07): US12 shipped — /explore parameter playground. 4 sliders (satellites, altitude, compute per satellite, ground stations) bound to a Svelte 5 runes state module ($lib/explore/params.svelte.ts). The scenario summary (total compute, bandwidth, coverage, round-trip latency, rough annual cost) recomputes live via $derived. Pure-math params.ts is split from the runes state so the math is TDD-testable in node env. 13 new TDD tests; pnpm test:unit 115/115 green. Live on https://localhost:8001/explore (55 KB, 4 sliders, 5-cell summary). Live charts (US11) and 3D orbit viz (US13) wire in next.
• v0.11 (2026-06-07): US8 shipped — $lib/components/charts/Chart.svelte is now the single point of ECharts integration. Dynamic-import in onMount, init, resize, dispose, reactive setOption (via Svelte 5 $effect), SSR placeholder, error fallback. LatencyChart.svelte (US24) refactored to a thin shell that builds the option and renders <Chart /> (51 lines → 24). Second consumer CompetitorTagChart.svelte renders a "competitors by tag" bar chart at the top of /competitors (pure-function data util $lib/utils/competitor-tags counts tags across the 11 profiles). 13 new TDD tests; pnpm test:unit 102/102 green. Sprint 2 closed (6/6).
• v0.10 (2026-06-07): US6 shipped — /competitors page reads the verified competitor profiles from 01-space-AI/01-research/competitor-profiles/ (default: ../01-research/competitor-profiles, env override: COMPETITOR_PROFILES_DIR) via a SvelteKit +page.server.ts. Renders 11 bento cards (10 verified + 1 unverified) with tag chips, summary, updated date, and a "View profile" link to the raw markdown. Verified badge is shown for profiles with a Verification & corrections section. New COMPETITOR_PROFILES_DIR env var + bind-mount in docker-compose.yml so the sibling research dir is reachable inside the container. 6 new TDD tests; pnpm test:unit 89/89 green; pnpm check 0/0; pnpm lint clean. Live on https://localhost:8001/competitors (~80 KB).
• v0.9 (2026-06-07): US5 + US24 shipped — /story long-form page with 5 curated narrative sections distilled from 01-research/ and 03-strategy/, and the US24 logarithmic-scale ECharts latency chart embedded in the latency section. Bugfix: landing-page CTAs (Explore, Story) were inert <button> elements with no href; made the Button component polymorphic (renders <a> when href is passed).
• v0.7 (2026-06-06): Foundation shipped. US4 added as highest-priority story and shipped: Traefik v3.6 reverse proxy in front of the SvelteKit container, TLS on :8001, HSTS max-age=31536000; includeSubDomains; preload, HTTP :80 returns 308 to https://localhost:8001/. Self-signed cert via scripts/setup-local-https.sh; Let's Encrypt cert resolver scaffolded for when a domain exists. The "Docker-only. HTTPS-only." hard-constraint callout now heads the Deployment section. US4 renumbered the future-sprint list (old US4–US20 → US5–US21). BUGFIX: Traefik router rule fixed from Host(localhost) || Host(127.0.0.1) to HostRegexp(^.+) so external-IP access (e.g. phone over VPN) works, not just localhost.
• v0.6 (2026-06-05): Expanded BACKLOG.md — split Sprint 1 into 3 smaller stories (scaffold, 3D earth hero, Docker), added Sprint 2-4 with US4-US16, added a cross-cutting section (US17-US20), added "Won't Do" non-goals, "Bug Fixes", "Stretch Goals", and a project-level Definition of Done. Added a story dependency map.
• v0.5 (2026-06-05): Locked the 4 remaining open questions: repo strategy (stay in monorepo for v1-v3), Globe.gl vs Three.js (Globe.gl for v1, Three.js escalation at v3), Svelte 5 (default), WebGPU (WebGL default, abstracted renderer). WebGPU note added to the tech stack table.
• v0.4 (2026-06-05): Added BACKLOG.md — active user-story backlog in JP's template format, SvelteKit-adapted. References added in the Related docs section. Also added the verification-pass skill to Related docs.
• v0.3 (2026-06-05): Wired the verification-pass workspace skill into the dev workflow — see "Source verification" section. Open question resolved.
• v0.2 (2026-06-05): Renamed folder 08-website/ → 07-website/ (Option A locked).
• v0.1 (2026-06-05): Initial draft.

📋 Template

Copy this template for new versions. Keep newest at the top (above the legacy section, below [Unreleased]). Omit any section that has no entries (don't add empty headers).

## [VERSION] - YYYY-MM-DD

### 🔭 Vision

_One-paragraph TL;DR: what is this version about? Why does it matter? What is the headline?_

### ✨ Added

- New feature 1
- New feature 2

### 🔧 Changed

- Change in existing functionality 1
- Change in existing functionality 2

### 🪛 Fixed

- Bug fix 1 (with brief context: what was broken, why, how it was fixed)
- Bug fix 2

### 💀 Deprecated

- Feature that will be removed in a future version

### 🗑️ Removed

- Feature that was removed this version

### 🔒 Security

- Vulnerability fix or hardening

### ⚠️ Known Issues

- Known issue 1 (link to BACKLOG US if applicable)
- Partial ships or in-flight work

### ✅ Verification

- Test count (X/Y unit + A/B e2e)
- `pnpm check` / `pnpm lint` / `pnpm build` status
- Live URL or other deployment evidence
- Commits: `abc1234` (feat: ...), `def5678` (fix: ...)

Conventions

• 📅 Versioning: Semantic Versioning — MAJOR for breaking changes, MINOR for new features, PATCH for bug fixes.
• 📅 Date format: YYYY-MM-DD (ISO 8601).
• ⬆️ Newest first: latest version at the top, oldest at the bottom.
• 🚧 [Unreleased] section: for in-flight work. On release, the unreleased entries get promoted to a dated ## [VERSION] - YYYY-MM-DD section above.
• ⚠️ YANKED marker: if a version is pulled from production, append [YANKED] to the heading and explain why in the body.
• 🔤 Emoji prefixes: keep these consistent (🔭 Vision · ✨ Added · 🔧 Changed · 🪛 Fixed · 💀 Deprecated · 🗑️ Removed · 🔒 Security · ⚠️ Known Issues · ✅ Verification).
• 🧹 No empty sections — only include a section if it has at least one entry. Better to omit than to add "none".
• 🔗 Link to BACKLOG USs: when a known issue maps to a US, link it ([US29](BACKLOG.md#us29-three-whys-narrative-deconstruction-why-this--why-now--why-us)).
• 📚 Legacy section: only the entries already there — never add to it. New entries always go above it, below [Unreleased].
• 🧾 Source of truth: the BACKLOG.md "Last Updated" line is canonical; the CHANGELOG entry is generated from the same data on each version bump (see BACKLOG.md § Process for the version-bump checklist).