- SequencePlayer: guard definition in bindStepsToDom; type SVG elements as
  SVGGraphicsElement so getBBox() resolves.
- Introduce shared StoryContent base interface; CategoryStory and
  SubcategoryStory extend it. CategoryStoryView accepts StoryContent so
  subcategory stories render without a type mismatch.
- StoryDiagramModal: containerEl is now $state and captured locally in the
  async render closure.
- SimulatorView: derive config from protocolId and load via $effect.
- ConceptTrigger: fold hover styling into enter/leave; add focus/blur so
  keyboard users get the concept tooltip.
- CategoryIcon: drive hover animation via CSS :hover (+ reduced-motion guard)
  instead of a JS handler on a roleless span.

npm run check: 0 errors, 0 warnings.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Genuine code fixes:
- Remove dead code: FOUNDATION_TEASERS/totalParts/foundationSections and an
  unused navigateToBookChapter import (DetailPanel); unused stubChapters helper
  (book/chapters); unused imports/vars in SearchBar, OutageView, NodeTooltip,
  SequencePlayer, layouts; vestigial gqlRequestLayer params.
- Drop the unused `color` prop from JourneyListView (it colors per-journey) and
  the matching caller arg; drop unused `stepIndex` prop from ActorStage + caller.
- Replace empty `interface MeshLink {}` with a type alias.
- Add keys to name-segment {#each} blocks (ProtocolHeader, MobileDetailSheet).
- Type the dev-only window globals (__dev/__tourDriver) in app.d.ts, removing
  three `as any` casts; make tourDriver const.
- Remove three unused eslint-disable directives in scripts/.

Rule-level config (with justifications in eslint.config.js):
- Off: svelte/no-navigation-without-resolve (app centralizes ${base} prefixing)
  and svelte/prefer-svelte-reactivity (all Map/Set are ephemeral locals).
- Allow _-prefixed unused identifiers.

Per-site disables with justification for intentional patterns:
- svelte/no-dom-manipulating at mermaid innerHTML injection sites.
- svelte/no-at-html-tags for trusted highlighted code + static SVG icons.

npm run check: 0/0. eslint: only no-useless-escape remains (handled next).
Build green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…en codemod

Adds scripts/fix-useless-escapes.ts, which consumes ESLint's JSON report and
deletes only the exact backslashes ESLint flagged as unnecessary (right-to-left
per line). A naive regex would have been wrong — e.g. in `Berkeley Unix\'s
\`.rhosts\``, the necessary `\'` (single-quoted string) must stay while the
redundant backtick escapes go; the codemod preserves exactly that distinction.

Verification:
- ESLint no-useless-escape: 537 → 0; 0 skipped (every position was a backslash).
- 34/35 changed files: esbuild-canonicalized output byte-identical → string
  values provably unchanged.
- text-parser.ts: two regex char-class simplifications ([\[ → [ inside classes),
  proven behaviorally identical across 23 edge-case inputs.
- npm run check: 0/0. Build green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Repo-wide `prettier --write` over source (Svelte/TS/JS/CSS/JSON). Pure
formatting — no code changed; check/lint/build all green afterward.

Also extends .prettierignore to exclude long-form prose working docs
(research/, podcast-blueprints/, *.md notes) where prettier's markdown
reformatting is non-idempotent and adds no value; those files are left exactly
as authored so the CI `prettier --check` gate stays meaningful for code.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- New ci.yml runs prettier/eslint (npm run lint), svelte-check (npm run check),
  and a production build on pull requests and non-main pushes.
- deploy.yml now runs the same lint + check gates before building, so main
  cannot deploy a red baseline.

(Test + cross-reference-validation gates are added in the test/validator
commits that follow.)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The earlier reactivity-warning fix wrapped simState.load() in an $effect, which
made Svelte flag unsafe state mutation at runtime (config/userValues written
inside an effect). The parent already remounts via {#key protocolId}, so the
load is genuinely once-per-mount; untrack(() => getSimulation(protocolId))
expresses that, silences the state_referenced_locally warning, and removes the
runtime errors. Verified clean in-browser across TCP/TLS simulate + stepping.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Unit (vitest, node env): 39 tests over the pure logic the content layer leans
on — sequence-parser (notes/arrows/blocks), text-parser (rich-markup grammar,
cross-ref resolution, bold-group recursion, strip), math, and color helpers.
Config added to vite.config.ts; scripts test:unit/test:unit:watch added; `npm
test` now runs unit then e2e.

E2e (playwright, against the production preview — no window.__dev): 15 tests
covering graph load + console-clean, /p/tcp detail + overview, Simulate tab,
404, category/subcategory, search, every registry/book/journey page prerender,
and a mobile bottom-sheet + no-horizontal-overflow check.

Deleted the demo scaffolding routes (src/routes/demo/**) and their placeholder
spec. CI (ci.yml) now runs test:unit + e2e.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…refs

Adds scripts/validate-cross-references.ts, wired into `npm run build` and CI.
It loads every registry and resolves all cross-references — structured
(protocol connections, categoryId, subcategory membership, journey steps,
comparison pairs, RFC/outage/pioneer/frontier protocol lists, story parents)
and inline [[id|label]] / {{concept}} markup in prose — failing the build on the
first broken link. Exports categoryStories / subcategoryStories / allPairs so it
can see those registries.

It immediately found 42 real bugs (refs that rendered as dead text / inert
tooltips); all fixed:
- Added 9 missing concepts: egp, congestion-collapse, ossification,
  tcp-handshake, http-verbs, comet, phishing, ttls, rsn.
- Corrected misnamed refs: {{as}}→autonomous-system, {{ct}}→certificate-
  transparency, [[head-of-line-blocking]]→{{…}} (it's a concept).
- Fixed protocol ids wrapped as concepts: {{ip|quic|websockets|mptcp}} → [[…]].
- [[ipv4]]→[[ip]]; [[ice|stun|turn]]→[[nat-traversal|…]]; [[datagram-transport]]
  (a subcategory, not a protocol) → plain text.
- Unwrapped 3 not-yet-added pioneers (Jay Kreps; Stanford-Clark; Nipper) to bold
  text rather than fabricate birth-year data — flagged to add later.
- pioneer dan-bernstein categories 'security' (not a category) → 'utilities'.

check/lint/unit/build all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CodeExample imported highlight.js core + 8 grammars statically (~150 KB raw /
~138 KB gz), so every first paint shipped it even when no code example was on
screen. Switched to a dynamic import() on mount: code renders as escaped plain
text until hljs resolves, then highlights reactively.

Measured on the prerendered / (single build, since chunk hashes are
non-deterministic): hljs grammar code is no longer in the initial chunk set;
initial JS is ~1171 KB gz across 23 chunks. mermaid and driver.js were already
dynamic and confirmed absent from initial. The remaining initial weight is
hand-authored prose — documented in docs/baselines as the next (riskier) win.

Added an e2e guard that http1's code example still highlights after the lazy
load. check/lint/16 e2e all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- static/screenshot.png (1.9 MB) was not referenced anywhere — removed.
- og-image.png was 2678×1612 / 1.9 MB, far larger than a social card needs.
  Resized to the standard 1200px width (1200×722) → 544 KB, a 71% cut. Format
  unchanged (crisp text, no JPEG artifacts) and no og:image dimensions are
  declared, so no meta changes needed.

Note: a spec-valid sitemap.xml needs absolute URLs, but no deploy domain
(CNAME/og:url) is configured in the repo — deferred until the canonical URL is
known rather than guessing the GitHub Pages path.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Added a global reduced-motion reset in layout.css that collapses animations,
transitions, and smooth scroll to near-instant when the OS requests it. This
covers the detail-panel slide-in plus the modal / tooltip / simulator-arrow
entrance animations that previously ignored the preference — in one place
rather than per component. The canvas graph already honors the same preference
in JS.

Added an e2e test asserting reduced-motion users still receive the detail-panel
content (animation off must not mean content hidden). check/lint/17 e2e green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

syncPositions copied simNodes[i] → nodes[i], which only works while d3-force's
internal array stays in creation order. A reordered or resized `nodes` array
would silently scatter every node to the wrong position. Now keyed by id.

Added simulation.test.ts (TDD): a reordered-array case that fails on the old
index logic and passes on the id match, plus an orphan-node case. Verified
in-browser that the graph still settles (live node positions populated, console
clean). 41 unit tests green.

Note: the render-loop perf micro-opts from the audit (memoize per-frame
node/connection maps, evict settled hover/dim animations, pool gradients,
delta-time viewport lerp) are deferred — they touch the canvas hot path with no
current user-facing bug, so they want a watched, profiled pass rather than this
autonomous run. Tracked in IMPROVEMENT_PLAN Phase 7.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The four diagram components (SequencePlayer, MermaidDiagram, StoryDiagramModal,
StoryDiagram) each repeated the dynamic import('mermaid') + initialize() with
the same base config (theme/security/font), differing only in their sequence/
flowchart tuning. Extracted loadMermaid(overrides) into mermaid-helpers so the
shared base lives in one place; each component passes only its own overrides.
Behavior unchanged — verified TCP sequence diagram + transport category-story
diagrams still render, console clean, build green.

The larger GenericLink (6 inline link components) and ModalShell (3 modals)
dedups from the audit are deferred: they carry visual-regression risk that
wants a watched, screenshot-diffed pass rather than this autonomous run.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

stripRichTextMarkup feeds plain-text surfaces (NodeTooltip body, AccessibleGraph
screen-reader text). A label-less ref like [[tcp]] resolved to '' there, silently
dropping the word. It now falls back to the display label (protocol abbreviation,
concept term, RFC number, …), mirroring how parseRichText renders the same ref —
so stripped text matches rendered text.

Currently latent (no bare refs sit in oneLiner/description fields today, and the
cross-ref validator can't catch it since the ref is valid) — this hardens the
path for future content. Added unit tests; 43 unit tests green, validate green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The cross-ref validator now flags duplicate ids. Protocol/pioneer/RFC/outage/
frontier collisions are fatal (none exist today, so this is a guard). Concepts
have 18 pre-existing collisions where two authored definitions share an id and
one silently shadows the other in conceptMap — reported as warnings for
editorial review rather than auto-deleting authored content or unilaterally
picking a "winner" (e.g. sip-invite, ns-record, bgp-update, and the company
entries apple/google/microsoft/meta/cloudflare each have two definitions).

Surfacing > guessing: the visible definition is valid today, so this is a
content-curation task for the owner, now tracked by the build.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>