feat: t27_spec brief by default; field lists and failure reasons from a .t27 spec (Closes gHashTag/t27#10090)

[gHashTag]

Closes gHashTag/t27#10090

Part of the epic gHashTag/t27#10066 (learnability: fewer tokens and round trips for cheap models). Stacked on #11 (t27-examples); the spec is gHashTag/t27#10092.

What

• t27_spec is brief by default: health, why it fails, loss, type-check errors, failing backends, size, summary, tags. fields: true returns the whole manifest record as before. For specs/demos/hello_world.t27 that is 619 bytes instead of 1330.
• t27_specs rows keep the short list they already had; it now comes from the spec.
• The hand-written brief() and failureReasons() are gone from server.mjs (35 lines out, 5 in). The field lists, the reason decision and the reason templates live in specs/t27mcp/spec_brief.t27, run as wasm/spec_brief.wasm with wasm/spec_brief.mjs; commands and sha256 are in wasm/BUILD. Built on the Railway lab.
• brief.mjs (19 lines) is glue: it passes the manifest counters to the wasm, copies the listed keys and fills the {} of the templates.
• smoke.mjs: the stale backendFailures check is fixed (see below) and two checks are added.

Hand-written JavaScript added: brief.mjs 19, server.mjs 5, smoke.mjs 9 = 33 lines (cap: 40 per file, 80 per PR). server.mjs shrinks by 30 net.

The smoke check that failed on #7

backendFailures tally matches the specs blaming a backend compared the sum of a per-backend count (737) with a count of specs (113). Not a server bug: the manifest's backendFailures counts specs per backend (c, js, rust, ts, verilog_hir, zig: 104 each; verilog: 113), and the sum of those is not a number of specs. The check now compares each backend's count with t27_specs {failedBackend} matched, and all seven agree.

Evidence

• npm run smoke: all checks pass, including t27_spec is brief by default, every failing row carries a reason from the spec (113 rows, no unfilled {}) and the per-backend check.
• The reasons from the wasm equal the JavaScript they replace on all 2402 manifest specs (113 failing, 0 differences).
• Spec side: 5 of 5 tests, 0 vacuous, negative controls (spec(t27mcp): spec_brief -- brief spec row fields and why a spec fails (Closes #10090) t27#10092).

🤖 Generated with Claude Code

[tri-claw]

Summary by t27-bees

The PR makes t27_spec return a brief record by default (health, failure reasons, loss, type-check errors, failing backends, size, summary, tags) instead of the full manifest, with fields: true for the old behavior. The brief field list, failure-reason logic, and reason templates move from hand-written JS in server.mjs to a .t27 spec compiled to WASM (wasm/spec_brief.wasm + generated wasm/spec_brief.mjs), with a tiny glue module brief.mjs. The smoke test's broken backend-failures tally is fixed to compare per-backend counts.

Worth a look

• brief.mjs:7 — new URL('./wasm/spec_brief.wasm', import.meta.url) assumes the WASM file sits next to the module at runtime. If the package is bundled or the file layout changes, the fetch will fail.
• brief.mjs:9–10 — The fill array indices (1=failedBackends, 2=tcErrors, 3=loss) must match the bit positions in wasm/spec_brief.mjs (W_BACKEND=2, W_TYPECHECK=4, W_LOSS=8) and the WHY_TEXT order. A spec change that reorders bits will silently produce wrong reason strings.
• brief.mjs:12 — The call W.why_mask(isFail, nodes, failedBackends.length, tcErrors, loss) must match the WASM function's parameter order exactly. A signature drift in the compiled spec will return a wrong bitmask with no error.
• server.mjs:348 — When fields: true, the response spreads the entire manifest row (...hit) and also adds whyFailing. The brief path (one(hit)) already includes whyFailing via cut(). If hit contains empty arrays (e.g., failedBackends: []), they appear in the full response but are filtered out by cut() in the brief response, creating an inconsistency.
• smoke.mjs:55–57 — The per-backend check calls t27_specs with limit: 1 but asserts n === m where m is matched. If matched returns the limited count (1) instead of the total, the check passes incorrectly. The test assumes matched is the total.

---

_{Read: 8 files, +95 -43, the whole diff. Head 351bc91. Model: openai-compatible/nvidia/nemotron-3-ultra-550b-a55b. Commands: @t27-bees review | summary | help | pause | resume. Free for every repository.}