[rhein1]

Implements the source-only Fable RRSI candidate and evaluation path for issue #24. A reviewed Fable commit is exported into a disposable repository with exact Git-blob provenance; only two finite-vocabulary prompt fragments can change. Candidate admission binds actual trees, modes, worktree bytes and independently tagged line edits while preserving the six workflows, ECF/evidence contracts and model/worker configuration.

The evaluator freezes source excerpts, factual assertions, uniform weights, split/model/runtime identities and retry/cost policy. It grades typed facts with exact evidence citations, retains missing slots and all reported retry consumption, and rejects unsafe, unknown-cost or incomplete comparisons. Held-out/OOD packets remain sealed. Live imports, acceptance and promotion stay disabled; synthetic fixtures do not establish improved Fable performance.

The repaired head is 3829e76c0072e140a4a8e0883ba56d8a570ef9c6. Six GitHub review findings and the additional nonterminal retry-consumption defect were reproduced against the prior source and repaired:

• All Git source reads disable replacement objects before resolving commits or blobs.
• Diff identity uses canonical unified bytes from verified ASCII blobs, independent of repository diff configuration.
• Aggregation keys bind task, split, trial and seed.
• Frozen assertions reject repeated fact paths as well as repeated IDs.
• Prompt screening caps each prompt at 1,024 characters and cumulative comparison work at four million character products before further comparison.
• Root and subcommand parser errors emit only source_operation_denied with exit 2.
• Every attempt must be terminal with known positive consumption for a successful retry to become comparable. Unresolved attempts retain reported consumption in totals and block comparison.

This change remains stacked on #25 at 3ab154be0dd72a6a1ef7967afbdb4b6b39297b00, preserving its Windows subprocess-cleanup and pipe-read classification fixes. Merge that prerequisite normally before retargeting this PR to actual main. Companion host budget/revocation/resume and reviewed Memory contracts: integrations #448. OpenShell lifecycle reconciliation: integrations #447.

Validation on the repaired candidate, Windows, Python 3.14.2 and Node 24.13.0:

• Source regression suite: 30 passed, 0 failed, 0 skipped.
• Evaluator regression suite: 28 passed, 0 failed, 0 skipped.
• npm test: 121 Node tests, 118 passed, 0 failed, 3 platform-specific skips. Its Laya Python suite reports 77 tests, 76 passed and one unavailable symlink test skipped.
• npm run validate, pwsh -NoProfile -File scripts/validate-package.ps1, npm run validate:artifact and git diff --check: passed. Packed artifact reran the complete tests and package validation in a disposable consumer; the owner installation was only a dry run.
• Independent read-only review of all six changed files and their relevant callers found no remaining source defect. This is not a formal approving GitHub review.
• Hosted validation on 3829e76: all eight checks passed (Ubuntu/macOS/Windows on Node 18/24, packed artifact, release gate). Literal tested merge ref bb5e71b has the same tracked tree as the reviewed branch head. Matrix runs have zero failures; platform-specific Node test skips are recorded in their logs. No required check was skipped.

Workflow trace: implementation and read-only review; parent owns edits/commit/push, worker /root/fable_ecf_review reviewed correctness, source identity, privacy and bounded evaluation. Recommended Luna worker compatibility fallback was declared. No recursive delegation.

Remaining qualification includes an authenticated host importer, durable all-in accounting and authority integration, private representative task suites, real isolated trials and owner-bound promotion/rollback. Required eligible GitHub approval and stack sequencing remain before merge. No paid model call, actual RRSI search, installed plugin change, Memory ingestion, OpenShell activation or deployment occurred.

[rhein1]

P2: Nonterminal retry history can be reported as cost-comparable

Reviewed head: 128a0eb373b1d39e8577f6331b00692b060357c2.

In aggregate_results, completion comes from the final retry, while earlier attempts need only positive reported cost/token values. An earlier unknown or timeout attempt with terminal:false, followed by a successful retry, returns comparable_synthetic:true, score 1.0, and zero missing trials. Its unresolved execution cannot establish complete final consumption or cleanup.

Expected gate: unresolved/nonterminal attempts must block cost comparability until their terminal outcomes and consumption are reconciled. Preserve all retry totals and the valid terminal failed-attempt followed by success control.

Reproduction: save this as a temporary Python file and run python -I -B PATH_TO_FILE from this head's repository root. It uses the actual evaluator and existing fixtures; the final assertion fails for both unknown and timeout, while the terminal-failure control passes.

import runpy
f = runpy.run_path("test/rrsi_evaluator_test.py")
e = f["e"]
config = f["config"]()
slot = f["slot"](config)
observed = {}
for status, terminal in (("failed", True), ("unknown", False), ("timeout", False)):
    rows = [f["observation"](slot, status=status, terminal=terminal, result=None),
            f["observation"](slot, attempt=1)]
    result = e.aggregate_results(observations=rows, expected_slots=[slot],
                                 config=config, synthetic=True)
    observed[status] = result["comparable_synthetic"]
    print(status, terminal, observed[status], result["total_cost_microusd"])
assert observed["failed"] is True
assert not any(observed[s] for s in ("unknown", "timeout")), observed

The existing evaluator suite passes 24/24 on this exact head. This is an offline synthetic-contract finding: accepted:false, promotion_allowed:false, and the live-import denial remain intact. No provider call, live activation, owner-file edit or push was performed for this review.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-01T23:24:46.942276Z	128a0eb	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 128a0eb373

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".

[jmborden1112-beep]

LGTM

Needs Fixing

None. RRSI source identity, bounded synthetic evidence and explicit pairing contracts remain consistent. The apparent single-arm aggregation concern was refuted: aggregation is per arm and match_arms handles pairing separately. No live optimization/promotion is implemented.

Requires Human Review

None beyond the repository approval gate fulfilled under the owner's second-account authorization. This is agent source review, not independent human sign-off.

Recommended Optional

None.

Create Follow-up Issue

None.

Workflow Trace: exact final head6e65019db917c52ee4c7f9fe36c0a02488215d3f includes merged25,22,23. Source lens /root/fable_stack_review plus root verification, gpt-6-luna medium fallback. Final combined local npm suite211passed/3platformskips/0failed; both package validators and diff check pass. Earlier RRSI source30/30 and evaluator28/28 Python tests passed. All eight final-head hosted checks pass, including Windows Node24, packed artifact and Release gate. No paid model trials, provider qualification, source improvement claims, publication, or runtime activation.