feat(experimental): Queue Auditor v0.1 → Human Review Workbench v0.5

[safal207]

Summary

Adds a five-layer, review-only control plane for the CML automatic Memory Proposal queue:

1. Queue Auditor v0.1 — measures queue pressure without collapsing Memory Pack identity.
2. Revalidation Planner v0.2 — recomputes canonical CML applicability + information quality + information fitness and preserves one decision per pack.
3. Live Revalidation Adapter v0.3 — separates immutable source identity from mutable PR/review/check evidence.
4. Semantic Acceptance Intake v0.4 — freezes one exact evidence-bound human-review packet per decision; ACCEPT / REJECT / DEFER remain semantic evidence, not repository authority.
5. Human Review Workbench v0.5 — enriches every frozen packet with generated lesson/situation/action plus source-path state on current main, then deterministically orders human review attention.

No layer grants merge, close, acceptance, execution, or policy-mutation authority.

Live queue

proposal_count                    = 36
stable_source_core_match_count    = 36
source_ancestor_of_main_count     = 36
Applicability.REVALIDATE          = 36
Quality.REVIEW                    = 36
Fitness.REVIEW_REQUIRED           = 36
semantic_acceptance_evidence      = NOT_COLLECTED

Full Memory Pack replay remains diagnostic rather than source identity:

source-checks mismatch  = 36
source-pr mismatch      = 8
source-reviews mismatch = 6
source-files mismatch   = 0
source-merge mismatch   = 0

Human Review Workbench v0.5

Priority is deliberately not a truth probability or universal score. It is lexicographic review ordering:

P0_SOURCE_SCOPE_MISSING
P1_SOURCE_SCOPE_DIVERGED
P2_REVIEW_CONTEXT_DRIFT
P3_OPERATIONAL_EVIDENCE_REFRESH
P4_CURRENT_SCOPE_MATCH

Within a class: lower generated lesson confidence → broader changed-path scope → older pack → PR number.

Exact-head live distribution:

P0_SOURCE_SCOPE_MISSING          = 0
P1_SOURCE_SCOPE_DIVERGED         = 20
P2_REVIEW_CONTEXT_DRIFT          = 5
P3_OPERATIONAL_EVIDENCE_REFRESH = 11
P4_CURRENT_SCOPE_MATCH           = 0

card_count                       = 36
pending_review_count             = 36
completed_review_count           = 0
authority_granted                = false

Top review cards are source PRs #234, #238, #264, #221 and #193 because their source scopes currently diverge from their exact source-merge snapshots. Queue rank does not assert semantic invalidity.

The path comparison is a net source-merge vs current-main state. A matching blob does not prove the path was never modified and reverted.

Exact-head evidence

Current exact head:

5b384c6e3161a6491b0be9fe2f7624b2ffa937eb

All product/evidence workflows on that SHA are green:

CI                                  31926850687  SUCCESS
Python Package Validation           31926850785  SUCCESS
Security Baseline                   31926850849  SUCCESS
Memory Queue Revalidation v0.3      31926850821  SUCCESS
Semantic Acceptance v0.4            31926850748  SUCCESS
Human Review Workbench v0.5         31926850828  SUCCESS

v0.5 exact-head evidence:

44 focused contracts = PASS
36 live review contexts = collected
36 workbench cards = verified
workbench digest = sha256:9c14a4ae2eb3ec11fba9a507bea91fd1cc7acb7409d2a384f98a41b6e088359b
artifact id       = 9258135071
artifact size     = 115642 bytes
artifact digest   = sha256:df9abbf849d8d1e3be28b53a1394e8999230a72bb5335a3dc442d55e956f0f16

Two housekeeping commits briefly created and then removed an empty placeholder during metadata handling. GitHub compare from the prior validated head 2cb30ea... to current 5b384c6e... reports files=[]; the final tree is content-equivalent, and the full exact-head workflow set above was rerun successfully on the new SHA.

Trust-root / bootstrap boundary

CML Trust Root Gate correctly remains FAIL-CLOSED on this exact head. Current status points to run 31926850470 and reports Protected CI contract rejected; inspect exact attempt.

This PR contains protected CI changes, including:

.github/trust-root/scripts/memory_queue_revalidation_collect.py
.github/trust-root/scripts/memory_review_workbench_collect.py
.github/workflows/memory-queue-revalidation.yml
.github/workflows/memory-queue-review-workbench.yml
.github/workflows/memory-queue-semantic-acceptance.yml

The protected base contract explicitly requires a dedicated bootstrap review for trust-root paths, approved workflow identities, or new workflows. This ordinary draft PR must not weaken or self-authorize that boundary.

Therefore this PR is not merge-ready even though v0.1–v0.5 product/review logic validates successfully. The protected workflow/collector changes need to be split or carried through the dedicated bootstrap process with explicit human + independent review.

Authority invariant

authority_granted = false
merge_authority = false
close_authority = false
acceptance_authority = false
execution_authority = false
policy_mutation_authority = false
review_completed = false

Machine evidence decides what must be inspected. The Workbench orders inspection only. Human judgment records meaning. Authority remains a separate transition.

This PR remains draft and unmerged.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 63b11e3d-7e53-47e7-8b0a-911a42781afa

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Relevant CML Memory

Deterministic retrieval for PR #287 at head 5b384c6e3161 against accepted memory on base 90c7fdaaf31a.

No publishable accepted memory matched this change

No publishable accepted memory met the relevance threshold.

---

Publishable candidates evaluated: 0 · selected: 0.
Public-repository policy surfaces only packs with visibility=public and contains_private_data=false.

CML memory is advisory: verify that the recorded constraints still apply. This comment grants no approval, execution, or merge authority.

[safal207]

Bootstrap path opened as #288: bootstrap(trust-root): approve Memory Queue v0.3–v0.5 CI identities.

#287 remains the ordinary product/evidence draft and must not be treated as authority for the protected collectors/workflows. The dedicated bootstrap candidate inherits the exact v0.1→v0.5 tree and adds only the five matching protected-file identities to the trust-root manifest.

Do not merge #287 before the bootstrap transition is separately reviewed and authorized.