fix(emitters): resolve rule references inside link text, and collapse a dead self-link (0.8.6)

[llima]

Summary

This PR bumps the version to 0.8.6. Merging publishes craftar@0.8.6 to npm once the release is approved in the npm environment.

What changes. 0.8.4 (AGENTS.md) and 0.8.5 (kiro) treated the text of a link they rewrote as opaque. A link whose text repeats its own target, [`.claude/rules/x.md`](.claude/rules/x.md), therefore kept a path the workspace may not have. This PR implements the approved spec rule references inside link text, the contract the user chose ("Token + colapsar"):

• In both agents-md and kiro, every rule reference inside a matched link's text is resolved like a token, with the same states and wording as the same mode.
• Collapse. If a link's own state reads "rule not in this workspace" and its text is exactly its own path (backticks kept), the link becomes the token form alone:
  • [`.claude/rules/cc.md`](.claude/rules/cc.md) → `cc (rule not in this workspace)`
• A link kiro keeps for claude-code (K2). Only the references in its text move. Non-references stay as written (spec Ruling 4).
• Warnings. A reference in a link's text is reported like a token and comes before the link's own entry.
• One code path. It is all in resolveRuleRefs (src/emitters/shared.ts): one token pattern and one collapse helper. agents-md.ts and kiro.ts do not change.

Which workspaces see update

The user approved the byte change at the spec's checkpoint, before the commit. They saw the before/after on a scratch Forge holding every example line of the spec, across six target sets.

workspace	at the first sync on 0.8.6
every workspace craftar import produced (rules targets: "*", claude-code a target)	no byte change in .kiro/, nor in AGENTS.md unless a link's text names one of its hand-written steering files (state B: it becomes .kiro/steering/<x>.md); a kiro link text citing an unknown name adds a kiro: warning entry
AGENTS.md written without claude-code, a link's text holds a .claude/rules/<x>.md reference	that text resolves (or the link collapses)
kiro, a link's text cites a rule kiro does not write	that text resolves (K2, K3 or D wording), or the link collapses
kiro, a link kept for claude-code whose text cites another name	only that reference moves

An affected workspace shows AGENTS.md or the .kiro/ files as update, and craftar sync --check exits 1 until it syncs.

Accepted odd forms

The user accepted these two forms as part of the contract:

• Repeated annotation. Non-exact link text in a dead link keeps the link's usual wording after its own resolved text: see cc (rule not in this workspace) (cc, rule not in this workspace).
• Fragment in the text. When the text carries a fragment it does not collapse: cc (rule not in this workspace)#part (cc, rule not in this workspace).

Tests

• New tests, written literally from the spec's examples. The 7 behaviour tests fail on 0.8.5. The 3 pins pass on both versions:
  • a Forge of the shape import produces keeps its AGENTS.md and kiro bytes;
  • reference-style links stay tokens;
  • links with a title stay tokens.
• Three kiro tests added during review, two of which fail on 0.8.5:
  • link-text-before-target order;
  • an import-shaped Forge where an unknown name in link text adds one warning entry;
  • an unknown name in a K2 link's text.
• No existing test changed. The only removed line is a missing final newline.

Known limits (unchanged)

• Nested brackets ([see [.claude/rules/cc.md](.claude/rules/cc.md)) match the outer [ (spec 15 §4.1). Recorded for the docs.
• A reference-style definition or a link with a title stays a token (spec 15 §12.6).

Test plan

• npm run typecheck: exit 0 on 77e395a.
• npm run build: exit 0.
• vitest without test/ci.test.ts (Linux): 853 passed / 5 skipped.
• npx tsx test/helpers/regen-golden.ts: no diff under test/golden/.
• Oracle: skipped. There is no fixture, and the user declined using a client workspace. Byte evidence:
  • goldens are unchanged;
  • the reviewer fuzzed 360,000 cases against 0.8.5: outputs differ only in bodies with a reference inside link text;
  • the whole-plan comparisons (an imported sandbox, and the spec 14, 15 and 17 Forges) are identical;
  • the approved before/after was re-run after every round, identical each time.
• Reviews: node-cli-reviewer 2 rounds, the second clean; docs-author 5 rounds on the README; the last finding (a steering name is state B in AGENTS.md) is fixed in 77e395a.
• CI green on this PR: 8 of 8 (ubuntu and windows × Node 22 and 24, runs 37307562016, 37307567200).