fix(assets): finish #4111 — dedupe/optimize the logo and add an asset-usage gate - #4139

…-usage gate

Refs #4111. That issue auto-closed on #4114's merge with two of its three
items undone. #4114 handled item 1 (five unreferenced favicons). This does
items 2 and 3.

Item 2 (the logo). Verified byte-identity with git hash-object/sha256: the
logo exists at exactly three tracked paths (not the issue's claimed four),
each ~1.39 MB — apps/viewer/public/logo.png, apps/landing/assets/logo.png,
docs/assets/logo.png. All three are genuinely necessary: apps/viewer is the
SPA at ifclite.com; apps/landing is a separate static site (its own
vercel.json, no build step) whose build-docs.sh also assembles the mkdocs
site under it for the ifclite.dev Vercel deploy; docs/ is additionally built
and deployed standalone to GitHub Pages by .github/workflows/docs.yml, which
requires docs/assets/logo.png to exist in the mkdocs source tree regardless
of the landing build. No safe dedup exists without adding a symlink or a
build-script copy step for marginal benefit, so all three are kept — but all
three are losslessly recompressed (Pillow optimize=True, verified
pixel-identical) from 1,457,769 to 1,320,603 bytes each, saving ~411 KB
across the three copies for zero visible change.

Also deleted four more unreferenced favicon originals under
apps/viewer/public (favicon-{16x16,32x32,48x48,64x64}.png, ~6.4 KB total) —
the same leftover pattern as #4114's five, flagged in the issue thread but
left for this PR. apps/viewer/public: 4,456,048 -> 4,312,468 bytes (26 -> 22
tracked files).

Item 3 (the gate). scripts/check-asset-usage.mjs fails when
apps/viewer/public gains a file no tracked text file anywhere in the repo
references (a substring search over basename / scan-relative path /
root-absolute path, deliberately permissive so it errs toward a missed dead
file rather than a live one flagged dead). Its ALLOWLIST covers
convention-fetched paths a browser or crawler requests by fixed name with no
in-repo link (favicon.ico, apple-touch-icon.png, robots.txt, sitemap.xml).
Detection logic lives in scripts/lib/asset-usage.mjs, tested against
synthetic asset/corpus lists in scripts/check-asset-usage.test.mjs (8 cases)
so a future change to the repo's real assets can't make the tests vacuously
pass.

Wired into .github/workflows/test.yml's node-tests job (both the gate and
its regression test), named check-asset-usage.mjs so check-test-wiring.mjs's
GATE_NAME_RE actually sees it — confirmed: check-test-wiring reports 50 gate
scripts / 98 scripts/ test files (up from 49/97) with this one counted as
wired, not listed among the @unwired-by-design exceptions.

Proved both directions locally: adding a git-added, unreferenced file under
apps/viewer/public makes the gate fail naming it; removing it makes the gate
pass again.

Not done: nothing — items 2 and 3 are both addressed here, item 2 by keeping
all three copies (justified above) rather than by deleting any.

#4111 is closed (auto-closed by #4114's merge), so this can't close it.
Either a maintainer reopens it, or applies `unqueued` themselves — a
self-applied label is stripped as SELF_APPLIED_LABEL, so this PR carries
neither.

Claude-Session: https://claude.ai/code/session_01QPHChk3Ve9N519A4kY7436

The gate's substring scan skipped .mts/.cts, so tools/demo-kit/derive-
variants.mts — which builds apps/viewer/public/samples/* paths — was
invisible to it. Nothing misfires today only because each sample name
also appears verbatim in apps/viewer/src/lib/tours/demo-kit.ts and
AGENTS.md, both already-scanned extensions; removing that redundant
mention while the .mts generator remained the only reference would
have made the gate call a live asset dead and block CI.

Moved TEXT_EXTENSIONS into scripts/lib/asset-usage.mjs so it's
importable from the test file without spawning the CLI. Did not add
.sh/.py/.rs: grepped every tracked file with those extensions against
every current apps/viewer/public asset's basename and found no
consumer, so there's nothing to justify the extra scan cost for them
yet.

Verified: constructed a case where an asset is referenced only from a
.mts file — 8/8-passing gate flagged it dead (RED), passes after the
fix (GREEN). Reverting the extension list fails the new regression
test (mutation check). Gate still passes for all 22 current files and
still fails for a genuinely unreferenced one. Runtime unaffected
(dominated by node/git-subprocess startup; two extra small files
read).

Refs #4111.

…bsolute-form test

check-ci-path-coverage.mjs flagged scripts/check-asset-usage.test.mjs for
reading vercel.json with no filter that can trigger the job. The literal
was a fixture label in a synthetic corpus entry, not a real read: the
gate's derivation only sees that it names a path that exists in the tree.

findUnreferencedAssets matches on the corpus entry's `content`, never its
`path`, so the label was free to change. Renamed it to a name with no
real-file collision instead of adding an allowlist exemption, since that
keeps the coverage gate able to catch an actual uncovered path in this
same file untouched.

…ment

The comment claimed .sh/.py/.rs were excluded because none of them
reference a path under apps/viewer/public. That's false for .rs: two
Rust tests (rust/processing/tests/instancing_dont_bake.rs and
rust/geometry/tests/clash_intersection_real_model.rs) reference
hello-wall.ifc and infra-bridge.ifc under that directory. The gate
still passes today only because both assets are also referenced from
McpPlayground.tsx, an already-covered extension.

Kept the exclusion rather than adding .rs to TEXT_EXTENSIONS: a
basename-only substring search over the repo's ~800 Rust files matches
generic test-fixture names (e.g. "manifest.json" in unrelated test
corpora) that have nothing to do with apps/viewer/public, so it would
add scan noise without closing the actual gap — the two real .rs
references are already covered elsewhere. Added a short comment at the
McpPlayground.tsx reference noting the gate depends on it.