coverage_release_0.13.0.md

Validated output coverage, 0.13.0

Version 0.13.0 fixes the 11 missed opportunities identified in the October 18 investigation. On the same original synthetic cohort, API participation increased from 10/32 (31.25%) to 21/32 (65.63%). All 20/20 reviewed safe opportunities reached line evaluation, compared with 9/20 before. The 11 original local-only controls still made no request. These are paired synthetic results, not a new installed workload participation rate.

Runtime changes

• Validate duplicate command envelopes only when text and structured content match the known schema exactly.
• Project homogeneous serialized command-result arrays, fulfilled Promise.allSettled arrays, and separately printed fulfilled records. Retain metadata, indexes, per-member boundaries, and the complete typed original. Validate every nonempty member's independent-line format before flattening. Carry its protected context into the projection.
• Keep the entire response when a member is rejected, unknown, mismatched, mixed media, structured, or connected prose. A later provider failure rolls back the whole batch, including originals and panel progress.
• Recognize exact passing Cargo test records and reviewed zero-valued diagnostic counters. Preserve other diagnostics on the same line, nonzero counters, malformed records, backtraces, and completion totals.
• Accept scalar-looking filenames only under verified path-list commands. Retain object and array guards. Recognize case-insensitive .log and numeric log rotations while keeping source and ambiguous suffixes exact.
• Recognize the complete read-only MCP action names list_commits, get_deployment_status, and lookup_merge_conflicts. Preserve credential and mutation guards and the saved generic MCP replacement policy.

The 95% classification gate, 5% omission cutoff, minimum-size default, provider selection, fallback setting, and session enablement behavior remain unchanged.

Live coverage results

The expanded coverage corpus contains 46 offline cases: 44 normal enabled outputs, one disabled control, and one deliberate later-batch provider fault. The live runner excludes deliberate faults and sends reviewed synthetic output only. Its 45-case run used the configured OpenAI provider with gpt-6-luna.

Expanded live cohort	Result
Normal enabled outputs	44
Distinct outputs forwarded to provider	24 (54.55%)
Completed line evaluations	23
Reviewed safe opportunities evaluated	23/23 (100%)
Local-only controls sent to provider	0/20
Provider requests	31
Required labels in prepared replacements	62
Required labels actually model-judged	17
Required labels protected before judgment	45
Proposed or prepared required-line losses	0
Saved originals verified	23/23
Prepared projected UTF-8 bytes removed	42,051
Reported provider input/output tokens	510,316 / 0
Called-output hook time, median / p95	831 / 2,206 ms

The original paired cohort holds fixture inputs and tasks fixed. Request expectations and projected line labels change for recovered adapters. All original typed-response wire byte lengths match; the full original transports are preserved for each prepared replacement. Projection changes mean that required-line counts are not directly comparable across the two versions.

Saved originals, hook replacements, and byte estimates do not acknowledge model-visible delivery. This run does not establish delivered context reduction or net cost savings. Zero loss on 17 model-judged required facts does not establish a general omission error rate.

Independent classifier holdout

Eight new reviewed cases cover all eight output classes through classification. They were not used to change the prompts, gate, or cutoff. Reports now separate normalized combined excerptable probability from the selected class's confidence. They include preset probability bins, confusion counts, Brier score, calibration error, and actual keep reasons.

The live provider selected the reviewed class in 6/8 cases. It kept all eight outputs: four true negatives and four false negatives relative to the reviewed excerptable labels. Six keeps had combined probability below 95%; two selected a non-excerptable class. No line evaluation or omission followed. Observed excerptability Brier score was 0.155825 and binned calibration error was 0.195.

This is a small holdout and does not establish population calibration. It confirms that unknown formats can still stop at classification after reaching the API. The release improves validated local recognition and leaves that remaining classifier limitation explicit.

Validation

• 288 Rust tests passed, with one ignored test. Formatting and Clippy passed.
• 169 Node tests passed, with three platform skips. All 16 Python checks passed.
• The seven-split offline matrix passed 618 invocations and 900 mock requests across both provider wire formats. Required-line losses and matrix failures were zero. Adapter faults verify whole-result rollback after an earlier successful relevance batch.
• Edge and Chromium smoke checks passed. Fifty Chromium layout rounds passed with ten toolbar-fallback checks and no page errors.
• Both release packages passed exact allowlisted-content, source-byte, version, and binary-target checks.
• The packaged VSIX passed two isolated native Windows launches with eight checks each. The checks cover exact hook installation, one registration, temporary-failure recovery, startup repair, restart idempotence, provider forwarding, and actual panel rendering on Codex 26.1007.21434.

The eval now records the exact bounded post-commit cleanup deadline message as a warning instead of inventing a hook error. It retains the warning in reports. Other stderr diagnostics and actual hook health errors still fail validation. The bounded runtime cleanup and recovery contract is unchanged.

Lab-Control could not allocate a container because its storage reserve was insufficient. Builds and checks used the existing locked WSL toolchain, two Rust build jobs, and normal user-local caches. Local reports remain under .local/quality/coverage-release-0.13.0/. paired-comparison.json contains the original-cohort comparison. live-classification/ retains the independent holdout results. Packages remain under .local/submission/.