Raise Advisor review limit to eight

[Cjbuilds]

Change summary

Fixes #19.

Raise the bounded global Advisor plan-review limit from five to eight while
preserving immediate PLAN_APPROVED exit and fail-closed behavior before
Executor work.

The native policy generator now has one authoritative
ADVISOR_REVIEW_LIMIT = 8. It derives the total-review cap, rounds 2–8 compact
ledger rule, and round-eight non-approval halt from that value. README, packaged
skill guidance, provider guidance, release qualification, and contract tests
are aligned. The packaged payload advances from 0.9.2 to 0.9.3 across every
required release-identity surface.

This replaces the unsafe stale work in PR #20; PR #20 is not merged or modified
by this change.

Validation

Exact reviewed head: 70eb5388f8ad5d72d2c0acc79e28fa1bd4f25379

• python3 scripts/preflight.py full --base-sha 8f0a2b1bcd65eb2b58b38679641555c83a639398
  — local git, compile, Ruff, focused tests, release identity, full tests, and
  real plugin lifecycle passed; hosted-only checks were intentionally skipped
  locally.
• python3 scripts/preflight.py portability --ci — every local portability
  module passed.
• python3 scripts/release_check.py --repo-root . --base-sha 8f0a2b1bcd65eb2b58b38679641555c83a639398 --head-sha 70eb5388f8ad5d72d2c0acc79e28fa1bd4f25379 --require-exact-shas
  — release metadata and identity valid for 0.9.3.
• Full deterministic suite: 384 tests passed, 5 skipped.
• Independent verification passed criteria A–G. Runtime monkeypatching the
  authoritative limit to seven changes all three generated bounds together,
  and the committed causal regression rejects any independently hard-coded
  eight-bound phrase.
• Hosted Python 3.11/3.13, macOS/Windows portability, and CodeQL remain
  authoritative and must pass on this exact SHA before merge.

Review attestation

{
"schema": 1,
"risk_tier": "security-state",
"repository": "Cjbuilds/Codex-Orchestration",
"base_branch": "main",
"reviewed_head_sha": "70eb5388f8ad5d72d2c0acc79e28fa1bd4f25379",
"reviewer_identity": "Codex independent verification worker Schrodinger",
"reviewer_route": "verification_worker using gpt-5.6-sol at high reasoning",
"threat_model": {
"assets": [
"Bounded independent plan review and Executor release authority",
"Compact findings-ledger integrity across stateless Advisor rounds",
"Plugin cache and release identity for the distributed 0.9.3 policy"
],
"threats": [
"A duplicated or hidden bound could make generated policy disagree about when review must stop",
"An unbounded or incorrectly extended loop could consume subscription capacity or release Executor without a valid approval state",
"Inconsistent documentation could cause users or release reviewers to rely on the obsolete five-review contract",
"Reusing a payload version could leave the deployed global policy on stale instructions"
],
"mitigations": [
"One authoritative integer drives every generated review-bound phrase and a causal monkeypatch regression proves the linkage",
"Rounds two through eight carry only the current plan and compact ledger, and round eight remains a fail-closed non-approval halt before Executor",
"Active guidance has a deterministic stale-cap sweep while unrelated five-based product facts are preserved",
"Release identity is synchronized at 0.9.3 and exact-base and exact-head release validation passed"
]
},
"negative_test_evidence": [
{
"category": "regression",
"evidence": "Focused tests fail on the 0.9.2 base, and the final causal test changes the authoritative limit to seven and requires all three generated bounds to change together."
},
{
"category": "negative",
"evidence": "A cross-surface sweep rejects stale five-review, round-five, and five-round cap phrases while explicitly preserving unrelated Fable 5, Opus 5, schema, effort, concurrency, and five-hour facts."
},
{
"category": "malformed",
"evidence": "Generated-policy regressions retain fail-closed rejection for stale plan versions, invalid or incomplete findings ledgers, required-route failures, and round-eight PLAN_REVISE without approval."
}
],
"findings_disposition": "Independent verification passed A through G. Its only non-blocking finding was missing causal constant coverage; that gap was repaired with an in-memory seven-limit regression and independently reverified with no unresolved findings on exact head 70eb538."
}