Doc/strictness cleanups from audit round 3

[lannbot]

Small doc/contract/strictness cleanups collected while auditing. None is a
correctness emergency; each is either a doc claim that no longer matches the
source, a contract that omits behavior the code ships, or a strictness gap
that lets a wrong shape pass silently.

Found in structured audit round 3 (2026-08-21). Audit/testing regime: #175.

README/docs

• Conformance total is stale. README.md:20 claims "1250 passing / 0
  failing"; the measured, expectation-pinned value is 1257
  (harness/browser/expectations/chromium.ts:54,
  harness/shell/expectations/*.ts). The README number predates the
  Close the non-pin conformance xfail residues (5 shim gaps + 6 named residues) #13/polymorph-tls: smoke the jco-leg surface under deltic (unexamined) #18 corpus prunes.

Embedder API

• Undocumented dual key spellings for world-level imports.
  runtime/src/embedder/instantiate.ts:499-514 resolves the kebab
  interface/resource id verbatim first, then falls back to camelCase;
  pick() accepts both pascalCase and raw kebab for resource classes
  (instantiate.ts:584, :646). contracts/embedder-api.md names only
  camelCase, and principle 3 permits liberal acceptance only where the TS
  type names a single canonical form. Consumers can (and will) couple to
  the unsanctioned kebab spelling. Either sanction both spellings in the
  contract or reject the non-canonical one.
• EmbedderOptions.verifyHash opt-out is not in the contract.
  runtime/src/embedder/instantiate.ts:129-131 (default true) lets a
  caller skip the sha-256 check that contracts/embedder-api.md A4
  states unconditionally ("fails loudly"). Contract-document the escape
  hatch or remove it.
• fromHost bool coerces instead of type-checking.
  runtime/src/embedder/values.ts:287-288 returns Boolean(v) for any
  input, while every sibling primitive lower rejects wrong-typed input
  with a site-named TypeError. "false" lowers as true; shape bugs
  pass silently.
• A dropped-but-unbound Stream handle can still be transferred live.
  runtime/src/embedder/streams.ts:293-299: drop() on an unbound
  handle only sets #dropped (there is no #host to drop). A later
  lower calls takeValue (streams.ts:216-227), which runs
  bindElement and never consults #dropped, minting a live shared end
  for the guest. takeValue should refuse a dropped handle.

Contracts

• contracts/plan-format.md exports[] schema is out of sync with the
  wire format. It omits the "instance" kind entirely — emitted by the
  shim since the resources corpus
  (crates/translator-shim/src/plan.rs:565), consumed at
  runtime/src/exec/executor.ts:851 and
  runtime/src/digest/digest.ts:209 — and documents the "type" kind
  with the wrong shape: the C2 amendment
  (contracts/plan-format.md:355-365) says
  {"kind":"type","resource":n}, while the wire shape is
  {kind:"type", name, type:{kind:"resource"|"value", …}}
  (runtime/src/plan/format.ts:259-278). The contract's own compat rule
  requires producer + consumer + contract in sync.
• Cache-key features component is inert caller-supplied salt.
  runtime/src/cache/core.ts:122-139 takes features: string[] = []
  from the caller and never derives it from the translator, while
  crates/translator-shim/src/lib.rs:111-113 documents
  producer.features as "part of the artifact-cache key" (echoed by
  docs/architecture.md §10 and contracts/plan-format.md:30-36). Sound
  today only because translatorBuildHash subsumes it (features are
  compiled into the shim wasm) — it becomes a cache-poisoning bug the day
  translateRaw grows a runtime features parameter. Doc drift plus a
  dormant trap.
• Digest name-flattening join is collision-capable and unspecified.
  runtime/src/digest/digest.ts:161-162 and :186 flatten with
  [...path, name].join("/"): import "a/b" with empty path collides
  with import "b" at path ["a"]. Not covered by
  contracts/digest.md's normalization rules. Preferred resolution is
  length-prefixing/nesting when the imports corpus lands — not blessing
  the join after the fact.
• The envelope container format has no contracts/*.md authority.
  WireEnvelope (runtime/src/plan/format.ts:284-295) is defined only by
  crates/translator-shim/README.md, yet it is the production deploy
  artifact (tools/translate output, embedder-api A4) and the cache
  determinism unit. It carries no version field of its own, and
  truncation/extension behavior is unstated.

WASI

• WASI WIT sets are not vendored in-repo. No wasi:io, wasi:cli,
  wasi:filesystem, wasi:random, or wasi:sockets WIT at @0.2
  exists anywhere in the tree (the only vendored wasi:* packages are
  clocks@0.3.0/0.3.1, http@0.3.1, sockets@0.3.0, and those sit
  under examples/guests/http-fetch/wit/deps/ as example-guest deps, not
  as an authority). Every 0.2-track semantic judgment therefore rests on
  unvendored authority, and source comments cite line numbers into WITs
  that cannot be checked in-tree (e.g. wasi/src/io.ts:69 cites
  "io.wit:23"). Same gap for wasi:cli/filesystem/random@0.3.
  Recommend vendoring the claimed sets — wasi/src/internal/fs_provider.ts:8
  claims WASI 0.2.12.
• header-error.forbidden is never produced.
  wasi/src/http.ts:324-337 and :351-365 validate syntax only;
  forbidden names are accepted into fields and then silently owned by
  fetch at send time. This is the platform-managed-headers divergence
  already recorded at wasi/src/http.ts:57-59 — the divergence record
  should name forbidden explicitly.
• Monotonic clock get-resolution claims 1 µs.
  wasi/src/clocks.ts:23-26 pins RESOLUTION_NS = 1_000n, optimistic
  where performance.now() is coarsened to 100 µs (non-cross-origin-
  isolated browser contexts).

Harness

• Deno conformance lane pins no totals.
  harness/tests/conformance_test.ts:120-139 fails on unexpected
  failures, stale xfails, and zero commands, but nothing bounds skip
  counts or asserts the pass count — so pass→skip drift is invisible on
  the primary lane (the shell/browser lanes do pin totals). Corollary: an
  xfail entry whose command degrades from failed to skipped is
  neither stale nor failing on the Deno lane.
• fileDeltas whole-file exemption has no stale detection.
  tools/browser/classify.ts:61,83-85: a fileDeltas entry turns ANY
  failure in the named file into an xfail, with no equivalent of the
  per-line stale check. Currently unused by all 9 overlay files, so this
  is latent machinery risk only; the first future use should get per-line
  deltas instead wherever feasible.
• Node openAt passes O_CREAT|O_DIRECTORY to the OS.
  wasi/src/filesystem_node.ts:236-262: for a nonexistent path with
  create + directory set, the ENOTDIR pre-check is skipped
  (existing === undefined) and the combination reaches open(2), which
  Linux documents as unspecified — it can create a regular file and then
  fail ENOTDIR, leaving a stray file behind. wasmtime rejects
  create+directory up front as invalid.