fix: read only the files a target emits as an ingredient's body (0.8.2)

[llima]

Summary

This PR bumps the version to 0.8.2. Merging it publishes craftar@0.8.2 to npm once the release run is approved in the npm environment. 0.8.1 is already published, so the merge order is satisfied.

What changes. Only an ingredient's body files are read for sections and {{param}} citations. A body file is a file a target emits and every emitting target renders as text. A file no target emits, such as a notes.md beside rule.md, is now ignored, and that has three consequences:

• A malformed marker in such a file no longer fails sync, status, diff, explain, ls or import, and no longer trips the schema: 1 gate.
• A {{key}} in such a file no longer makes forge unify or import refuse.
• take: param and take: section on such a file are now refused with … is not emitted by any target.

No emitted byte changes. The README has an Upgrading → to 0.8.2 note, which covers plans saved under 0.8.1.

How a declared file is matched. emittedFile and bodyFile take the ingredient directory, which is required (string | null). They match a declared file/files entry by path.join(dir, …), the same call the emitters use to read it. As a result, every file spelling that 0.8.1 synced still syncs with the same bytes: ./rule.md, /rule.md, ../r/rule.md, a/../rule.md, a script's ./run.sh, and so on. A table test pins this against what 0.8.1 does, as measured on Linux. null is passed only by the two importer sites, where the importer builds canonical names itself.

Errors. When a body file lies outside the ingredient directory, sits behind a symlinked directory, or is spelled with a different letter case on a case-insensitive file system, the error is now a user-facing message naming the file. Before, it was an internal: error.

Each dir call site is covered by a test that fails when its directory is replaced by null. The exception is ctx.text in sync.ts, where the mutation is equivalent: it receives the declared string as written.

Disclosures

• An existing test was rewritten. In d5e41ad, the U1 test "refuses a variant-only file that would bring markers in…" in test/unify.test.ts was rewritten to use a dir-layout skill, because a rule's notes.md is no longer a body file. The reviewer judged the rewrite faithful to the original intent. A new test pins the relaxed rule-notes.md behaviour (ddcb02a).
• Two commits are red in isolation. 8e2f9a5 and baf1faa each fail 37 tests in test/unify.test.ts. cdf49c8 restores green, and every later commit is green on its own.
• Some commit texts are imprecise. History was not amended.
  • The body of 0daeb10 says the regression "broke 0.8.1 syncs". The regression came from d5e41ad; 0.8.1 itself was fine.
  • The subject of af1968d over-promises.
  • The body of 1e40309 claims parity with path.join. That parity only came with defdaf5.
  • cdf49c8 and 26c583f are docs: commits that also change source: extract.ts fallbacks, and the error message in sync.ts.
  • The subject of 02b6c98 says "every" site; the remaining sites are covered by 1b14bf1, 53ade25 and cb946d9.
• Out of scope, to be recorded as tech debt. The script and hook emitters build output paths from the raw declared string, for example .claude/scripts/./run.sh. 0.8.1 does the same, and this PR does not change it.

Test plan

• npm run typecheck: exit 0 on fc0b41c
• npm run build: exit 0
• vitest without test/ci.test.ts (Linux): 790 passed / 5 skipped
• Oracle: skipped, no fixture available. The reviewer's spelling probe compared plan() output between 0228d69 and HEAD on 32 cases, and the paths, bytes, warnings and errors were identical.
• test/golden/** is unchanged.
• Windows: not run locally; CI ran windows-latest on node 22 and 24, and both passed. The skipIf(win32) symlink test ran only on Linux.
• CI is green on this PR: ubuntu and windows × node 22 and 24, in runs 37221319156 and 37221341611.