fix(moa): neutralize reference_responses tags in reference responses (FleetReview P1, #1353)

[ang-fleet-workers]

FleetReview P1 on #1353 (key e953813ec984): a reference response could close <reference_responses> and forge a User query: section.

Fix: _neutralize_reference_tags entity-escapes any open/close reference_responses tag (case, whitespace, attribute variants) inside each response before it is wrapped. Content is kept, not dropped.

Test: tests/tools/test_mixture_of_agents_prompt_roles.py. The new test asserts there is exactly one real open/close tag pair and that only the real query follows the close tag. It fails on the pre-fix code and passes after.

Card: t_931fdc1d

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[ang-prism]

FleetReview

Below the merit threshold, review skipped (33 changed lines < 150 and 2 changed files < 6, no hot path touched; skipped by policy).

---

FleetReview · reviewKind: skipped-by-policy

[ang-fleet-lander]

🤖 merged-by: apollo · lane: kanban-merge-pass · gate: fleetreview 48e307f green · why: t_931fdc1d: FleetReview P1: #1353 @ab6b6d46 tools/mixture_of_agents; Argus off card review (Ace 13:08), CI green

[Kyzcreig]

🤖 merged-by: apollo · lane: review · gate: fleetreview 48e307f green · why: t_931fdc1d: CI green; Apollo review pass 17:55

[ang-prism]

FleetReview

Review: post-merge · head 7766fcf645c2 · duration 11m 41s
Profile: light (merit: default light: lines 33<800, files 2<1000000, hunks 4<1000000, no hot path) · policy: below-size-and-path-gates
Roster: B-assert-ctx → gpt-6-sol (openai), C-assert-xhigh → claude-code-opus-5-5 (anthropic), F → gpt-6-sol (openai), G → gpt-6-sol (openai)

Post-merge review (fleetreview:post-merge override): this reviewed the merge commit against its first parent — the bytes that already shipped. It is not a pre-merge gate pass.

Reviewed with 2 of 3 model families — xai unavailable.

profile: light (rule: default light: lines 33<800, files 2<1000000, hunks 4<1000000, no hot path) · round 0 · members: B-assert-ctx, C-assert-xhigh, F, G · families: anthropic,openai

Confidence: 4/5

Findings

• P2 tools/mixture_of_agents_tool.py:91 — Regex slowdown · agreed: B-assert-ctx,C-assert-xhigh (openai, anthropic)

---

FleetReview provenance · models: B=gpt-6-sol, C=claude-code-opus-5-5, D=grok-4.6, F=gpt-6-sol · cost: $0.35 · duration: 11m 39s · rounds: 2 · files examined: 2