[echoomegaprime]

Purpose

Add one isolated SQLite regression for the combined PR #40/#41 candidate (PR #49). Start two recovery workers at a thread barrier against a historical FAILED dispatch/QUEUED run/BOUND reservation and assert a single terminal run, released reservation, zero remaining metered runs, and exactly one negative compensation event.

Evidence

• The same test logic was executed on FORGE's isolated combined-source scratch checkout: 1 passed in 11.69s, exit 0 (Node/Nexus diagnostic 2026-11-11 02:21 UTC); source SHA-256 of scratch test 9092ff64181be2470735d817cc24cdea90177f005bb09a62bfd48ab1056f7ff3 before documentation-only header changes.
• Exact GitHub commit 3dcd40bd1db8694b5843e034c0333f1d1e9b0924 was read back. Compare against PR fix(subscriber): combine transaction rollback and exhausted dispatch recovery (#40 #41) #49 head 0ab617ab6ed97e1a40e5cb3d160215dc93dc86d5: one new test file, 43 additions, zero deletions.
• Await exact-head hosted CI and repeat this test on the committed blob. The scratch pass is not hosted CI and not a full concurrency stress certification.

Safety / rollout

• Draft PR targets PR fix(subscriber): combine transaction rollback and exhausted dispatch recovery (#40 #41) #49's feature branch, not main, to avoid changing the signed release candidate or competing with other owners.
• Test-only; no production database, service, secrets, permissions or protected approval operations touched.
• Merge only after independent review and CI, followed by renewed exact-head certification and backup/rollback verification for the eventual integrated candidate.

[echoomegaprime]

2026-11-11 02:35 UTC — independent exact-head CI and source review

Fresh GitHub readback confirms this draft PR's exact head 3dcd40bd1db8694b5843e034c0333f1d1e9b0924 (one new test file, blob 13fa345e78f6737ae880bd56260971970e5e1448). Hosted CI run 42945718482 is completed/success, job test #130277914721 completed/success. Individually verified successful steps: exact checkout, install, compile, full deterministic suite, P1 verification, P6 deployment enforcement, P7 subscriber governance, evidence upload. Previously recorded exact-source isolated rerun: 1 passed, exit 0.

Independent test review: the test uses a two-thread barrier against one historical FAILED/QUEUED/BOUND fixture and asserts INFRASTRUCTURE_FAILURE, release of the bound reservation, certification_runs meter zero and exactly one negative dispatch-crash:% usage row. That verifies one important concurrency interleaving; it is not a repeated stress test, crash-at-commit test, or deployment attestation. Suggested follow-ups: repeated interleaving or multi-process stress under SQLite busy timeout; failure injection between state transition/reservation release/compensation to verify atomic rollback; audit entry uniqueness and redaction; bounded reconciliation of many orphan rows.

No integration into PR #49, merge, service change or production data mutation occurred in this review. Preserve independent review, backups and signed production E2E as release gates.