fix(website): a half-printed module was coloured as a whole one

[gHashTag]

The Spec Explorer's health chips read 976 working · 100 warnings · 338 broken.

Most of the 338 was one defect in one place. gen-js and gen-ts emitted references to
names they had not bound, so 126 specs produced modules that are syntactically perfect and
throw a ReferenceError at import — one unbound name costs every declaration in the file.
gHashTag/t27#4528 fixes that (gHashTag/t27#4529). This PR vendors the compiler that carries
the fix and teaches the catalog the third answer it now gets back.

The third answer

A declarations backend can produce an artifact that is real and incomplete: it prints
what it can and names what it could not, inside the artifact, as data (__NOT_EMITTED__, a
frozen list). Whole and half-printed must not share a colour — a partial module is exactly
what a "Working" chip would be lying about.

t27-corpus.mjs reads notEmitted from the backend rather than regexing
__NOT_EMITTED__ back out of the code — which would be a second, weaker implementation of
something the backend already knows exactly. It records partialBackends, tags the spec
issue/backend-partial, and puts it in warn.

The catalog spec stops holding numbers nothing checks

specs/catalog/onboarding.t27 gains HEALTH_FAIL_UNPARSED and HEALTH_PARTIAL, both
derived from the manifest by onboarding-from-spec.mjs and therefore checked by
check:onboarding.

Two things were removed rather than updated:

• The hand-typed class breakdown. It added to 154 beside a sentence that said 145 — a
  breakdown nothing checked, which is the one kind of number this document is not allowed to
  hold.
• assert HEALTH_FAIL_JS_ONLY > 0. True the day it was written, and a lie the moment
  the class it counts is fixed. A test that forbids its own subject from reaching zero is a
  test against the work.

Measured over the corpus this PR ships

1419 specs — five more than the last sync, picked up from upstream on the way past.

	before	after
✓ working	976	1017
⚠ warnings	100	184
✕ broken	338	218
js backend fails	335	211
ts backend fails	335	211

211 is exactly the set that produces no AST at all — c, rust and zig fail the same
211, and no backend can serve them. The js/ts-only failing family is zero. The remaining
7 are specs that parse and then lose a backend (6 × verilog_hir, 1 × verilog).

71 specs now carry a non-empty __NOT_EMITTED__ and are counted apart — neither whole
nor broken.

Cache

The wasm is 731 581 bytes, so its content tag moves from 3f1e073c963c86e2 to
bb39b9a5ca412a26. No browser holding the old immutable entry is ever asked for it — the
lesson of #1078, and the reason the tag is derived from the bytes rather than stored.

Gates run locally

check:spec-catalog (1419 exact source links) · check:onboarding (spec tests 9, asserts
49, all hold) · check:t27-evolution · check:explorer-languages · check:discovery ·
check:shared-core — all PASS. audit:en / audit:ru need CHROME_PATH and did not run
here; they run in CI.

Sequenced after gHashTag/t27#4529 in the sense that the source of the vendored wasm lands
there, but this PR is self-contained: the wasm bytes are in it.

🤖 Generated with Claude Code

{
  "version": 1,
  "head_sha": "a41600ad7f921412e64a47e3c4185b7dcc487743",
  "summary": "The Spec Explorer called 338 specs broken. With the js/ts value-layer fix vendored from t27, the manifest reports 218, and 71 specs that emit a real but incomplete artifact are counted apart as partial instead of being called either whole or broken.",
  "changes": [
    "t27-corpus.mjs derives partialBackends and notEmitted from the compiler's own targets.js.notEmitted field rather than regexing __NOT_EMITTED__ back out of the emitted code.",
    "A spec whose backends all emit, but where one of them omitted declarations, is now health warn and tagged issue/backend-partial instead of counting as fully working.",
    "notEmitted across backends is combined with a maximum, not a sum: one declaration refused by both js and ts is one missing declaration, not two.",
    "onboarding.t27 gained HEALTH_FAIL_UNPARSED and HEALTH_PARTIAL, both derived from the manifest by onboarding-from-spec.mjs so neither number can be typed by hand.",
    "A prose sentence asserting a count that nothing checked was replaced by an explanation of the class plus a constant the manifest verifies.",
    "Regenerated the vendored compiler wasm, manifest.json, shared-core.json, universe-atlas.json, agents.t27 and llms.txt against the new corpus of 1419 specs."
  ],
  "tests": [
    {
      "command": "npm run check:onboarding",
      "result": "reports the generated documents are up to date",
      "status": "passed",
      "evidence": "iterated sync, regenerate, sync until the printed constants stopped moving at corpus size 1419"
    },
    {
      "command": "npm run check:spec-catalog",
      "result": "the compiler is pinned at the new content hash",
      "status": "passed",
      "evidence": "prints compiler pinned at ?v=bb39b9a5ca412a26, derived by t27-wasm-tag from the new wasm bytes"
    },
    {
      "command": "node apps/website/scripts/t27-corpus.mjs",
      "result": "manifest health counts 1017 ok, 184 warn and 218 fail",
      "status": "passed",
      "evidence": "the same corpus read 976 ok, 100 warn and 338 fail before the vendored compiler changed"
    },
    {
      "command": "cargo test --bin t27c (upstream gHashTag/t27, branch fix/js-ts-value-layer)",
      "result": "1736 tests passed with none failing",
      "status": "passed",
      "evidence": "1736 passed, 0 failed, 2 ignored; the artifact vendored here is built from that branch"
    }
  ],
  "limitations": [
    "The 218 specs still counted broken produce no AST at all, so no backend can serve them; that parser work is deliberately not in this change.",
    "These health numbers only move on the live site once the upstream t27 compiler change lands and the wasm vendored here is rebuilt from it.",
    "Seven specs parse and then lose exactly one backend, six verilog_hir and one verilog, and this branch does not examine them.",
    "Test results are reported from local runs on this machine and are not independently rerun by the report generator."
  ],
  "tags": ["t27", "SpecExplorer", "compiler", "codegen"],
  "blog": {
    "title": "Why 338 specs looked broken when only 218 were",
    "summary": "A declarations backend that had no spelling for one initialiser used to refuse the entire module, so the Spec Explorer painted healthy specs red. Naming the omission inside the artifact instead turns them green, and introduces a third answer sitting between working and broken.",
    "outline": [
      "The health chip on app.t27.ai counted 338 broken specs, and that number turned out to describe how the js and ts backends reported failure rather than how many files were damaged.",
      "A declarations-only backend has a third possible answer: an artifact that is real but incomplete, printing every declaration it can and naming in __NOT_EMITTED__ each one it could not.",
      "Per-backend omission counts are combined with a maximum rather than a sum, because a single declaration refused by both js and ts is one missing declaration and not two of them.",
      "The five compiled backends receive null for that field rather than zero, so a catalog reading the manifest can tell no such question apart from no omissions at all.",
      "Every printed constant is derived from the manifest, so a corpus that grows by five specs in the middle of landing cannot leave a stale hand-typed number behind it."
    ]
  }
}