Add --adversarial falsification persona to the document bounce

[alanshurafa]

What

Adds an opt-in --adversarial reviewer persona to co-evolve-bouncer.sh. Until now the reviewer seat ran a one-line "be adversarial" role regardless of stakes. This gives the document pipeline a structured falsification method — premise challenging, assumption surfacing, decision stress-testing, simplification pressure, alternative blindness — vendored as a repo-owned template so the pipeline stays free of any plugin runtime dependency.

Two usage modes:

• Cross-AI bounce (default): --adversarial runs the persona on the claude reviewer seat, codex composes the resolution.
• Internal same-model: --adversarial --agents claude,claude runs it single-vendor, with a logged warning that same-model loses cross-vendor disagreement.

Design

• Persona is orthogonal to run mode — no new mode value (the scorer's baseline logic keys off mode).
• Composes with existing controls: --lens becomes a focus line appended to the persona; --chain swaps only the pass-1 critique stage (chain-critique-adversarial.md); defend/tighten untouched.
• New additive reviewer_persona field (light/adversarial/lens) in state.json, via an optional 8th arg to init_bounce_state defaulting to light — other callers unaffected, scorer ignores it.
• Adaptation from the compound-engineering adversarial-document-reviewer: dropped the peer-persona deferrals (no peer personas exist in this single-seat pipeline), emit [CONTESTED]/[CLARIFY] markers instead of JSON, kept depth calibration with per-pass finding caps to limit marker flooding.

Files

• templates/co-evolve/role-reviewer-adversarial.md, chain-critique-adversarial.md — new persona templates
• co-evolve-bouncer.sh — flag, template validation, usage(), build_reviewer_preamble branch, chain swap, REVIEWER_PERSONA + banner + same-model warning
• lib/co-evolution.sh — optional reviewer_persona in init_bounce_state
• evals/BOUNCE-RUNNER-CONTRACT.md, skills/co-evolution/SKILL.md, README.md, CLAUDE.md — docs
• tests/adversarial-persona-simulation.sh — new 6-scenario sim

Verification

• tests/run-all.sh: 33/33 suites (incl. new sim at 6/6); bounce-scorer regression 7/7.
• Dogfooded live end-to-end on a seeded-flaw plan: cross-AI bounce reached convergence_status: converged with reviewer_persona: adversarial. Head-to-head against the compound-engineering persona (same model) matched its coverage of all 7 seeded flaws.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 73d17ce304

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve the lens focus in chain adversarial mode

When users combine --adversarial --chain --lens <name>, the chain critique prompt is built directly from chain-critique-adversarial.md, so the lens-specific focus line added by build_reviewer_preamble is skipped. That makes the documented composition silently ignore the requested lens in the high-stakes chain mode, e.g. --lens security auditor produces a generic falsification prompt instead of a security-focused review.

Useful? React with 👍 / 👎.