Complete unchanged persisted run decoding reuse
RE-VERIFY: This document is a checkpoint. Recheck the remote branch head and test summaries before continuing. Production was not changed in this lane.
Design decisions in effect
2026-10-23 owner requirements:
- "Keep that contract: still fingerprint, just skip the decode when the fingerprint matches."
- "If you believe that test must change, do NOT weaken it. Make the change in its own commit and stop to report it in your final file."
- "Never raise timeouts, cut iterations, or add JVM flags. Never merge, enqueue, publish or deploy."
Remaining work
Review the fingerprint-test correction, then implement bounded immutable decoded base reuse with current source freshness, publication ordering, bounded retention, close/restart, and concurrent-reader coverage. The required invariant table and candidate public scenarios are in findings.md. The cache itself is not implemented at this checkpoint.
Evidence and mechanism
loadRunFromDisk always calls buildRunFromPersistedSource. Each call constructs JSONObject and runs the complete record decoder. The supplied public-read scenario checks returned fields and unchanged bytes, then counts complete decodes across 32 more reads. Production evidence identified 71 samples in persisted decoding and a separate 69 in snapshot replacement; this lane addresses only nonresident reads.
HOSTCPU evidence: https://github.com/CodexCoder21Organization/DigitalOceanDropletWui/tree/wip/laneHOSTCPU-evidence/evidence/HOSTCPU
Supplied reproducer: https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/laneHOSTCPU-record-read
Test measurement issue
The allocation test uses onPersistedRunRecordDecoded to count its claimed fingerprint work. The observer is documented as a complete-decode counter. The exact 10,000-count assertion therefore requires repeated full decoding. A separate count-only fingerprint observation is needed to preserve the fingerprint count and allocation limit while allowing complete decodes to be skipped. The stop boundary above applies after that correction.
Relevant refs
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/laneFIXREC-evidence
| Branch | Code commit | State |
|---|---|---|
| https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/laneFIXREC-evidence | https://github.com/CodexCoder21Organization/BuildTestEmbedded/commit/5deb85c0ddf9d12eb0cbe0f1ba067118d884813a | Unmerged; no PR; later evidence-only commits on the same branch |
Results: initial unchanged-read reproducer red on attempt 1, all 3/3 attempts red at 32 additional full decodes. Corrected allocation test passes on its first attempt (1/1), exact 10,000 fingerprint observations and unchanged 160 MiB ceiling. Summaries and XML identities are in this directory. The fingerprint test uses the new onPersistedRunSourceFingerprinted observation after real source digest computation; onPersistedRunRecordDecoded retains its original meaning. No cache reuse, fix reversal, complete suite, or PR was completed because the brief requires stopping after this test correction.
Continue only after the orchestrator reviews the separate test correction. The complete suite remains required once the cache fix and its row scenarios are ready. Do not carry the old receipt ordinal, skip current lease/live-count sidecars, or suppress invalid archive-metadata diagnostics through reuse.
No code in this lane was deployed or published. No merge or queue entry was made.
Resume commands
Fresh clone under the new lane directory, fetch the evidence branch and rebase on main. Read README.md, DocumentationRepository architecture/TESTING.md, and evidence/FIXREC/findings.md. Targeted runs use scripts/test.bash --local --test <name> --log <file>. The one complete run uses --remote, following the original brief. Read attempt-history in XML rather than treating a runner pass after another attempt as a first-attempt pass.
<!-- FIXREC2-CHECKPOINT-START -->
FIXREC2 checkpoint at 11:57 UTC, 2026-10-23.
Fixed source: https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/FIXREC2-decoded-records at f231c8d9b00cda83a5e81b43738944e853b64194. The approved fingerprint observer commit remains included.
Targeted run: 8/8 pass on attempt 1, including the original 10,000-fingerprint and 160 MiB allocation assertions. Reversal control: all seven scenarios red on attempt 1 and all 21 attempts red: https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/FIXREC2-reversal-control.
The single complete run is still active locally after remote health preflight returned CONCURRENT_HANDLER_LIMIT at 256 handlers. Local suite slot was acquired around 09:18 UTC. Session 40044, parent process 13790; log workspace/FIXREC2/complete-suite-local.log and expected XML workspace/FIXREC2/complete-suite-local.xml. XML is not yet present. Source head has stayed unchanged throughout this run.
The lane has reached roughly four hours elapsed. The brief also says never to stop a lane while its complete suite is running. User clarification on these two rules is pending; the existing run continues without changes. No PR yet: final full-suite counts, latest-main rebase, final targeted rerun and PR creation remain. Do not merge, enqueue, publish, deploy or monitor CI after PR creation.
Findings: out/rFIXREC2-findings.md. All passing source work and control evidence are already on remote branches.
<!-- FIXREC2-CHECKPOINT-END -->