feat(decay): add LLM-assisted relevance recheck at maturity thresholds

[foundatron]

Closes #11

Changes

1. tentacle/llm/prompts.py — Update decay check prompts

• Change DECAY_CHECK_SYSTEM output schema from {"still_relevant": bool, "reasoning": "..."} to {"action": "halt"|"decay"|"accelerate", "reasoning": "...", "comment": "..."}
• halt = still relevant, skip decay; decay = proceed with normal -1; accelerate = no longer relevant, drop to 1
• comment field is optional, used for GitHub issue comments on halt/accelerate
• DECAY_CHECK_USER template: no structural changes needed, just ensure the action vocabulary is documented in the system prompt

2. tentacle/decay.py — Add LLM recheck at threshold crossings

• Expand apply_decay signature: apply_decay(store: Store, *, grace_days: int = 30, interval_days: int = 60, llm_client: LLMClient | None = None, context: str = "", repo: str = "", model: str = "claude-haiku-4-5", dry_run: bool = False) -> int
• After computing new_maturity but before applying, check if the decay crosses a threshold (4→3 or 2→1)
• At thresholds, call _llm_recheck() which:
  • Retrieves issue body via store.get_analysis_by_id(issue.analysis_id) — uses suggested_body (or "(no body available)" if None)
  • Calls llm_client.complete() with decay check prompts using the configured model
  • Parses JSON response; on parse failure or unknown action, returns a default "decay" action (fail-open)
  • Returns a RecheckResult (NamedTuple: action: Literal["halt", "decay", "accelerate"], reasoning: str, comment: str)
• Action handling:
  • halt: Skip decay, log DecayEntry with reason "llm_recheck:halt — {reasoning}", call comment_on_issue() if comment is non-empty
  • decay: Proceed with normal -1
  • accelerate: Set new_maturity = 1, log DecayEntry with reason "llm_recheck:accelerate — {reasoning}", call close_issue() with comment, call store.update_issue_status(issue_id, "closed")
• After any decay (threshold or not) that results in maturity 1, call close_issue() + store.update_issue_status()
• Catch BudgetExceededError inside the per-issue loop: log warning, set llm_client = None for remaining iterations (mechanical decay only)
• If llm_client is None, skip LLM recheck entirely (backward compat)
• Return count of decayed issues (unchanged semantics — halt doesn't count, accelerate counts as 1)

3. tentacle/issues.py — Add close_issue() and comment_on_issue()

• comment_on_issue(issue_number: int, comment: str, *, repo: str, dry_run: bool = False) -> bool
  • subprocess.run(["gh", "issue", "comment", str(issue_number), "--repo", repo, "--body", comment], ...) — list args, no shell
  • Returns True on success, False on failure; logs warning on failure; fail-open
• close_issue(issue_number: int, comment: str, *, repo: str, dry_run: bool = False) -> bool
  • subprocess.run(["gh", "issue", "close", str(issue_number), "--repo", repo, "--comment", comment], ...)
  • Same error handling pattern
  • Note: verify gh issue close --comment support; if unavailable, fall back to comment_on_issue() + close_issue() without --comment

4. tentacle/db.py — Add get_analysis_by_id

• get_analysis_by_id(self, analysis_id: int) -> Analysis | None — simple SELECT * FROM analyses WHERE id = ?, reuses existing _row_to_analysis converter

5. tentacle/cli.py — Wire LLM client into cmd_review_backlog

• In cmd_review_backlog:
  • Create CostTracker and LLMClient (with config.scan_budget, monthly budget, etc.)
  • Fetch context via existing fetch_context() pattern (same as cmd_run)
  • Pass llm_client, context, repo=config.target_repo, model=config.decay_model, dry_run=args.dry_run to apply_decay()
• Add --dry-run flag to review-backlog subparser

6. tentacle/config.py — Add decay_model config field

• Add decay_model: str = "claude-haiku-4-5" to Config dataclass (alongside existing filter_model and analyze_model)
• Add to _apply_toml mapping and DEFAULT_CONFIG_TEMPLATE

Review Findings

• Errors: 1
• Warnings: 6
• Nits: 4
• Assessment: NEEDS CHANGES

The most critical issue is #7 — repo defaulting to "" means any call to apply_decay without an explicit repo argument (which is the case when no LLM client is configured, but mechanical decay still reaches maturity 1) will attempt gh issue close --repo "" and fail. The issue will be stuck open at maturity 1. Either make repo required when close behavior is possible, or guard the close_issue call on repo being non-empty.