fix(locomo): v8 — twenty-word answers + first-step invalid-completion retry

[fazpu]

What

Two answer-stage fixes measured on the conv-47 v7 re-score (91/150), shipped together as the v8 protocol identity:

1. Answer-length rule: six → twenty words. The six-word cap caused 7 outright failures and made 19 misses structurally unwinnable (gold answers are enumerations longer than six words; the agent's compressed half-answers get judged wrong). New rule: shortest phrase that fully names the requested entities/values, at most twenty words, no explanations. Guard + message updated in the runner; tests pin 20 passes / 21 fails on both sides.
2. First-step invalid-completion retry. Under v7's larger prompts (mean 22.8K tokens_in/question), luna emits non-JSON on the first agent step for ~15% of questions (23/150), which the fix(locomo): bounded retry on invalid reader completions + protocol-pinned reasoning effort #178 retry deliberately did not cover — those questions died with zero retries. A unified invalid_completion_attempts counter now gates the same bounded budget for first-step and reader-position failures alike; reader_attempts semantics unchanged (validator-compatible); new optional first_step_retries / total_first_step_retries fields (defaults keep old states parsing); provider outages remain non-retried.

Protocol: full-v8/full-v8-strong, adapter version bumped, tool catalog hash unchanged, locked fingerprints updated, v7 error strings migrated. Docs: design changelog, decisions.md D78 record, benchmarks corpus next-steps note.

Expected effect on conv-47: mechanical losses (30 items ≈ 20%) recovered → ceiling ≈ 115–120/150 vs 91 under v7. Rerun follows merge.

Reviews

• Grok-4.5: APPROVE WITH NOTES — verified budget non-escape (first-step retries count against per-question, global, and cost budgets), reader-position regression intact, old-state parse compatibility, fingerprint/identity coherence incl. error strings; notes are stale secondary docs (runbook, overall_design index), non-blocking.

• Claude (final review): verified unified retry gate, validator compatibility, 20-word guard both sides, ran benchmarks suite (77 passed) + ruff format --check.

• I have read and agree to the RememberStack Contributor License Agreement v1.0.

[github-actions]

Coverage report

This PR does not seem to contain any modification to coverable code.