add prior-analysis search to work-on-task-assistant Phase 6

[bborbe]

Problem

Phase 6's three semantic searches are all artifact-type-scoped:

1. search_related("{task_title}")
2. search_related("{task_title} runbook procedure")
3. search_related("{task_title} guide")

They find runbooks and guides. A prior decision — recorded on a task or KB page — matches none of those framings, so it stays invisible even when indexed and highly relevant.

Observed failure (2026-08-24)

A task to deploy 12 new per-symbol candle handlers passed Phase 6 clean and shipped a merged PR before anyone noticed [[Review candle-command-handler per-symbol design]] — a five-week-old analysis that had already identified that exact fleet as the cluster's top pod-count offender and decided to consolidate it.

The page was indexed, scored 0.55 on a topic query, and was sitting on the same day's daily note. All three queries missed it because it is neither a runbook nor a guide. The conflict only surfaced when the human objected on capacity grounds — reaching, manually, the conclusion the vault already held.

Change

• Adds query 4: search_related("{task_title} prior analysis decision design", top_k=5)
• Hits ≥0.5 whose conclusion constrains the task (a decision to consolidate / deprecate / freeze / reduce what the task would grow) are reported under a distinct ⚠️ CONFLICTING PRIOR WORK heading — deliberately not folded into the guides list, since a guide says how to proceed while this says whether to proceed at all
• Self-check updated: all four queries required; a clean guides result is not a pass if query 4 never ran

Docs-only change to the agent definition. make precommit PASS.

[ben-s-pull-request-reviewer]

{
  "verdict": "request-changes",
  "summary": "The PR correctly specifies Phase 6 query-4 behavior (all four queries, distinct ⚠️ CONFLICTING PRIOR WORK heading) in the agent definition and the CHANGELOG is properly updated under Unreleased. However, the success_criteria enforcement at line 452 is stated as a requirement in prose — it is not backed by any runner/wrapper implementation in scope — meaning an agent that skips query 4 would still pass Phase 6 without violating anything in this PR.",
  "comments": [
    {
      "file": "agents/work-on-task-assistant.md",
      "line": 452,
      "severity": "critical",
      "message": "Phase 6 enforcement (all four queries required, query 4 included) is stated in success_criteria but has no implementation. An agent that skips query 4 would still produce a \"clean guides\" result and pass Phase 6 — contradicting the self-check text. The calling runner/wrapper must track that query 4 was actually invoked, or this requirement will never be enforced."
    },
    {
      "file": "agents/work-on-task-assistant.md",
      "line": 452,
      "severity": "major",
      "message": "No regression test covers Phase 6 query-4 enforcement. If the runner is later updated to validate query-4 execution, there is no test asserting that a missing query 4 produces a FAIL verdict — the new behavior has no automated guard."
    }
  ],
  "concerns_addressed": [
    "correctness: Phase 6 self-check enforcement — addressed: flagged as critical (no runner implementation)",
    "correctness: Query 4 conflict results reported under ⚠️ CONFLICTING PRIOR WORK heading — not an issue: confirmed unambiguous in text",
    "correctness: 'A clean guides result is not a pass if query 4 never ran' condition — addressed: same as first concern",
    "tests: No integration/smoke tests for Phase 6 query-4 behavior — addressed: flagged as major"
  ]
}