# Exact final-head evidence index

PR [#20093](https://git.cafe/pingdotgg/t3code/pulls/20093) ([GitHub mirror](https://github.com/pingdotgg/t3code/pull/20093)). Baseline [968dd51704](https://github.com/pingdotgg/t3code/commit/968dd51704441138559a3fe9c348719f2b11e0c6) → final [22aaa53e02](https://github.com/pingdotgg/t3code/commit/22aaa53e02e86bb3728abbe8432fb15c910487ae). Every selected after-side scenario was freshly run at exact22aaa, including natural601s expiry, clock boundaries, all failures, batching, native core, mutations, reconnect/restart and compatibility harness. Unchanged before968 captures are reused; corrected HTTP/pagination and new closed-only/open-index scenarios are paired fresh runs. Old72 after traces are rehearsals, excluded. No historical process is relabelled.

**Corrected gaps:** E2 final P503/P503/P200, failed envelopes data:null with SERVICE_UNAVAILABLE errors, forced new-list retry then recovery and cached-success read; D survives. O4/G1 B2/version202 absent from both first pages and actually delivered by both continuations, ending with no cursor. First page limit1 is per repository; continuation explicitly uses limit99 with returned cursors. Baseline first-page fallback includes B1; final page1 has A only. This is not same-size paging proof. Assertions over raw HTTP/RPC pass. Negative controls reject old503-success-data and empty continuation. New closed-only B7→open B1 scenario proves direct CLOSED/OPEN summaries, omission on final after manual invalidation and explicit B push, and recovery after index catches plus fresh read. Immediate invalidation does not guarantee immediate GitHub indexing; warm positive coverage has the intended ten-minute tradeoff.

54 selected captures;3280 fixture HTTP attempts;1553 list/continuation/provider-list deliveries plus4 direct summaries=1557 normalized deliveries. New exact final core110 HTTP=109 GraphQL+1 REST against retained baseline215 HTTP=214 GraphQL+1 REST. Isolated core109→51 HTTP,108→50 GraphQL. No overlapping scenarios are summed as savings. All live rateLimit.cost fields are null. Simulated fixture cost sums are separate: core214→110, not live quota or inferred estimates.

Core workload traces preserve3 genuine Codex turns/commits/pushes with checkpoints and receiver counts2/3/4; separate3 explicit app pushes and1 no-op. First-PR C and closed/open B each have separate explicit scenario pushes. Other timing/variant/mutation scenarios have their own pushes; seed pushes precede capture. No forge branch push/comment/body edit/merge/browser/delegation/auth configuration/mode/MCP changes. Worktree clean; all owned server PIDs exited; scratch about1GB, below10GB.

What else changed: from72 to22aaa only removal of false empty-repository source GraphQL comment; every other tracked blob, including generated query and executable runtime, identical. Both source exports verified against2482 tracked Git blobs per pin. This equivalence is supplemental provenance; final claims still use exact22aaa processes. Fixture corrections and evidence checks only; core native workload unchanged. No tracked code edits by this evidence agent.

## Published artifacts and exact local paths

| Artifact | Public URL | Exact local path | Bytes | SHA256 |
|---|---|---|---:|---|
| driver-397cfbfc41e0860c.tar.gz | [driver-397cfbfc41e0860c.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9850ec9f9d67ad62/driver-397cfbfc41e0860c.tar.gz) | `/tmp/takeover-20093/correction/artifacts/driver-397cfbfc41e0860c.tar.gz` | 22043 | `397cfbfc41e0860cb5f5d32b840337f3e1056aa06b5d3180b036b18a0a05d7c3` |
| transcripts-fdbb1fd3349b8096.tar.gz | [transcripts-fdbb1fd3349b8096.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f52013bd6dd6c80b/transcripts-fdbb1fd3349b8096.tar.gz) | `/tmp/takeover-20093/correction/artifacts/transcripts-fdbb1fd3349b8096.tar.gz` | 1145701 | `fdbb1fd3349b8096315b249cde835cb806a54d3d04eecfae64b6f33c79d7190f` |
| matrix.md | [matrix.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/13312845d44b3947/matrix.md) | `/tmp/takeover-20093/correction/publish/matrix.md` | 17733 | `31908e6a12a5dd681bc6a2c2c698ce0ece712dab720988984cb0f7b2728d593a` |
| counts.md | [counts.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0f92ff9630d06767/counts.md) | `/tmp/takeover-20093/correction/publish/counts.md` | 4122 | `cf5fbd490723a00aa930bfa83e2474f3284f60eff8a7e7e5038f6fabe3609ede` |
| manifest.json | [manifest.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d0090c123215688d/manifest.json) | `/tmp/takeover-20093/correction/publish/manifest.json` | 50693 | `db7fd5debb0c7814cd06f291918108ec396830d2f174da920105b79960c35046` |
| request-ledger.ndjson | [request-ledger.ndjson](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e000577bce44c24f/request-ledger.ndjson) | `/tmp/takeover-20093/correction/publish/request-ledger.ndjson` | 6796038 | `1a3997e09ae4037efcad2da03f37ec5b7102a5b943c7a68910fde99d3b8197f2` |
| deliveries.ndjson | [deliveries.ndjson](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9529a2f62751d0de/deliveries.ndjson) | `/tmp/takeover-20093/correction/publish/deliveries.ndjson` | 1892482 | `9f4a00fd724b2a233b30144bef7a92314ff1a9858fdb2814e089bb09b7cf9d17` |
| ledger-notes.md | [ledger-notes.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/277da6155f77cb77/ledger-notes.md) | `/tmp/takeover-20093/correction/publish/ledger-notes.md` | 4455 | `ba3850d3ffc6d673caaac21c5c0711886c76aba075caebb61092429b8e2a6971` |
| assertions.json | [assertions.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d00690f09dc281bd/assertions.json) | `/tmp/takeover-20093/correction/publish/assertions.json` | 2369 | `912cb86abf607607e78d4abc90d4e02b14cd046530b93788dad6b09c6cb9cc40` |
| assert-all-after.json | [assert-all-after.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4f073382806372e7/assert-all-after.json) | `/tmp/takeover-20093/correction/publish/assert-all-after.json` | 4135 | `cb1ed927e8ae0fda250c47092644049d7ff27e1890e469f7ddf093c378a800a3` |
| negative-controls.json | [negative-controls.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7a53fdaa356cc26c/negative-controls.json) | `/tmp/takeover-20093/correction/publish/negative-controls.json` | 388 | `26db0472a3bcbfcb8282615823204399a605fe82866c2cbb3cfddf74b28bc0ec` |
| capture-map.json | [capture-map.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ae0e0dafb8069a34/capture-map.json) | `/tmp/takeover-20093/correction/publish/capture-map.json` | 5177 | `d90e688dadd5cdeb3e963306688082b544d489d2970741ef678284aaa6b31bdf` |

Raw selected captures are `/tmp/takeover-20093/correction/publish/captures/<capture>/{http.ndjson,rpc.ndjson,fixture-state.json,server.log,source-pin.json}`; retained baseline lacks historic source-pin.json and uses audited original archive provenance. Reusable driver is `/tmp/takeover-20093/correction/driver`; complete SHA/blob provenance is manifest.json plus source-export-verification.ndjson inside transcripts. capture-map.json maps all old72 after names to new final22aaa captures. Raw selected files are hashed individually in manifest. Source exports/homes/node_modules/secrets/public-config are excluded from both archives.

## Replay and independent audit handoff

Extract hashed driver archive to a fresh scratch root with driver/ under it; README contains explicit git-archive exports, installed-dependency links, synthetic-repo setup, exact normal-RPC commands and expected assertions. Extract transcript archive separately. A fresh auditor should verify both corrected rows against raw records and independently rebuild at least one row from the NEW driver archive (prefer final-corrected-http or final-corrected-pagination-v3). Run assert-evidence.py on that fresh replay. Parent must commission this audit; no delegation performed here. This index is evidence, not merge approval.

I also extracted the newly built transcript archive and re-ran HTTP/pagination evidence assertions on its contents successfully. All seven corrected/new/core row checks,23 additional after-scenario checks and two negative controls pass. These are evidence validation, not product tests.

Unverified/limits: no live GitHub quota/index timing measured; fixtures simulate HTTP truth and index lag. No live GHES auth/Forgejo CLI integration; isolated service/provider harness is labelled. Real Codex only; unused provider integration remains static-path evidence. Exact equal-time TestClock case is timer-first only. P3 create-PR/combined action busy-window Cartesian product is not separately captured. Pagination continuation page size changes intentionally. Fresh independent proof audit and independent replay remain pending.

FRICTION

| category | happened | prevention |
|---|---|---|
| fixture semantics | Old503 had usable coverage and old continuation delivered no row. | Corrected captures plus raw response/count/identity checks and failing negative controls. |
| baseline paging | Baseline fallback put initial proposed next row on page1. | Added boundary prefix and older B2; assert B2 absence then presence on both pins. |
| source provenance | Comment-only final SHA arrived after old captures. | Reran all selected after-side scenarios at22aaa; retain unchanged baseline968 only. |
| evidence checker path | First all-after check looked for expiry controller log under sequential naming. | Corrected checker to the actual separate expiry log; reran all checks successfully. |

Produced with gpt-6.1-sol, Codex harness in T3 Code. Upload skill: [ghdrop](/home/ubuntu/.codex/skills/ghdrop/SKILL.md).