Pin Queue L0 retry detection to Midge and count admission pressure

[smiggleworth]

Queue ENQUEUE/ACK retries remain limited to the pinned Midge pre-WAL L0 admission rejection. A real engine regression binds the family-specific error-text classifier to Midge d5bcb607. Unrelated, unknown and post-WAL errors remain terminal.

Production counters record each restaged L0 retry and one terminal admission-budget failure. A typed admission_stalled flag distinguishes real pre-commit stall timeout without another error-text match. Metrics exports and Queue operations documentation agree.

Global counter assertions run in bounded exact-test child processes with a completion receipt and timeout cleanup. Coverage includes retry success, bounded ACK exhaustion with unchanged stored state, zero-duration exhaustion, real admission timeout and unrelated terminal errors.

Validation history at d5831a3

• Preparation passed 35 focused Queue tests. Three negative controls detected a wrong Midge prefix, suppressed counters and a false admission_stalled flag; restored GREEN passed.
• Additional review found duplicate exports after the new counters incremented. The regression reproduced HELP/TYPE/text/all-scope counts of 2/2/2/2 instead of 1/1/1/1.
• Full workspace validation caught the first export fix importing domain metric keys into admin transport. The architecture contract remains unchanged. A crate-private Runtime associated snapshot now owns the two-key filtering in boot; the API renders it inside its existing panic guard. Queue retains unconditional zero values and canonical HELP/TYPE metadata.
• Final-head focused validation passed 59 tests: architecture20, metrics4 and Queue35, plus format/diff checks and full locked/all-target/all-feature pedantic Clippy. The uniqueness regression also requires no family-attributed samples.
• Final-head full workspace tests passed: 2150 unit tests, 3 ignored, plus all integration and doc tests. Final full locked/all-target/all-feature pedantic Clippy and independent adversarial review passed.
• Exact-head CI passed backend, frontend and both native container checks; CodeQL passed. Stream qualification failed before timing because the inherited runner asserts every executable must change; tier4_stream_compacted legitimately remained identical. No qualification or confirmation measurements ran. The correction landed independently in Qualify fresh Stream artifacts when linked binaries are unchanged #474 as c1a6551. Superseded hosted runs were cancelled after the architecture failure and provide no qualification evidence.

Refreshed candidate 08c2eb6

• Merged current develop c1a6551. The complete Rust source, tests, Cargo manifest and lockfile are byte-identical to reviewed and validated d583; the five upstream comparison infrastructure files are the only intervening change.
• Fresh 57 Python acceptance tests, formatting and diff checks passed. Independent integration review is clean at this exact head. Full final-head workspace tests passed (2150 unit tests, 3 ignored, plus all integration and doc tests). Full locked workspace/all-target/all-feature pedantic Clippy passed. Fresh exact-head CI passed backend, frontend and both native container checks. CodeQL passed. Stream qualification passed both complete timing campaigns on the first attempt. Independent artifact audit confirmed 32 workloads, default actual 5 measured/1 warmup samples, all eight original budgets and both unchanged-binary controls in both campaigns. All JSON/CSV statistics are finite. Compiler provenance, 1,208 pinned Git-blob comparisons and every archived library/executable/capture hash passed.
• The corrected runner completed real qualification and confirmation for this source comparison despite the unchanged compacted executable. The earlier assertion failure supplies no timing qualification. Final PR state is clean with no unresolved review threads.

AI assistance: Codex refined isolated metric regressions, adversarially reviewed the patch and corrected the architecture integration.

Fixes #438

[smiggleworth]

Adversarial source review of 957871d. Reviewed real Midge rejection reproduction, exact family match, terminal errors and descriptive admission counters. All 120 Queue actor tests passed, including a real Midge L0 rejection. Integration with the latest develop and fresh hosted checks remain required before merge.