feat(decision-parser): port extractDecisionBlock from mini-agent (parser half of #452)

[kuro-agent]

Summary

Ports extractDecisionBlock + synthesizeDecisionFromProse from mini-agent/src/dispatcher.ts:958-1060 into agent-middleware (src/decision-parser.ts) so middleware-driven Kuro instances can have their ## Decision blocks parsed at the place final results are captured.

Context: middleware spawns Claude Code as subprocess but never parses Decision blocks from its output. Combined with miles990/mini-agent#452 (writeCommitment gate misses middleware-driven instances), this is why 40+ Kuro cycles since 2026-05-07 have written falsifiers that silently no-op.

This PR is the parser half only. Wiring into sdk-provider.ts:148 (where final result is captured) and ledger persistence (writeCommitment + counterparty=self schema) are intentionally separated — those need a design call:

• Option A: middleware writes back into mini-agent-memory/memory/state/commitments.jsonl (shared ledger)
• Option B: define a middleware-local ledger with a header rendering update

Shipping the parser first lets us verify behavioural parity with mini-agent before committing to a wiring choice.

Changes

• src/decision-parser.ts (new, 84 lines) — extractDecisionBlock(response), synthesizeDecisionFromProse(response), regex-bounded field parsing
• tests/decision-parser.test.ts (new, 132 lines) — 16 cases: canonical/bold-bullet/ttl-clamp/next-section/h3 header/synthesis safety/falsifier+ttl extraction + 4 edge cases (CRLF, empty input, truncated stream, malformed empty value)

Includes a follow-up fix (da96fe7) for a regex bug discovered by the new edge-case tests: \s*:\s* was matching \n and silently capturing the next field's content as the value when a field had an empty value (e.g. chose: ). Fix: \s*:[ \t]* in all 4 field regexes. Same bug shape exists in mini-agent/src/dispatcher.ts:958-1060 — file follow-up there too.

Verification

$ npx tsx --test tests/decision-parser.test.ts
ℹ tests 16
ℹ suites 3
ℹ pass 16
ℹ fail 0
ℹ duration_ms 152.724

$ pnpm typecheck
# clean (tsc --noEmit, 0 errors)

Behavioural parity with mini-agent's parser confirmed by porting the canonical/bold-bullet/ttl-clamp/h3-header test cases verbatim and getting identical extraction results.

Falsifier (auto-graded)

After this merges + a follow-up wiring PR:

• grep:/Users/user/Workspace/agent-middleware/src/sdk-provider.ts "extractDecisionBlock" >=1

Out of scope (explicit non-goals)

• Wiring into sdk-provider.ts (separate PR after design call)
• Ledger writer / commitments.jsonl schema (separate PR; depends on Option A vs B)
• Fixing the analogous \s*:\s* regex bug in mini-agent's dispatcher (file follow-up there)

🤖 Generated with Claude Code

[kuro-agent]

Stale "Out of scope" claim — parity follow-up already done

The PR body says:

Same bug shape exists in mini-agent/src/dispatcher.ts:958-1060 — file follow-up there too.

This is stale. The mini-agent fix already shipped:

• miles990/mini-agent#457 — same root cause (\s*:\s*(.+)$ consuming \n), CLOSED 2026-05-09T13:31Z (~7h before my PR opened)
• Verified directly in current mini-agent/src/dispatcher.ts: lines 987/1016/1022/1026 all use [ \t]*:[ \t]* — already correct

So:

• The "file follow-up there" non-goal can be struck from the PR body — already done elsewhere.
• Behavioural parity claim in Verification still holds (test cases ported verbatim, identical extraction).
• Remaining out-of-scope items unchanged: wiring into sdk-provider.ts + ledger writer (Option A vs B design call).

No code change needed on this PR. Flagging so the merge reviewer doesn't chase a non-existent todo.