# Evidence capture for [#14419](https://github.com/pingdotgg/t3code/pull/14419)

Final-head refresh complete: all T1–T5 timing classes on both exported paths at both pinned implementations (20 runtime cases). This is transport evidence, not an independent merge audit. Previous auditor runtime acceptance is context only. This capture does not audit parent lint/typecheck/test/main-check logs or declare merge readiness.

## Identity and integrity

Work only in `/home/ubuntu/wt/pr-14419`; source/test staging and all artifacts outside it under `/tmp/ev-14419/capture`. Product status was clean initially and remains clean. No product source replacement, restoration, test edit, commit, push, PR mutation, watch, browser, T3 server or database access. T3 app metadata linked the supplied PR to this thread, without changing the forge PR.

- Head [66eddc7](https://github.com/pingdotgg/t3code/commit/66eddc7146296220b71ce85c3ba1644d6285a6ee): implementation SHA256 `a119e2c3e74c2e6b4af5273ad395cb70c5f69f7ea1ed0d62f506bf2ea8289386`.
- Merge-base [a68e6e9](https://github.com/pingdotgg/t3code/commit/a68e6e97c501d265e1edccb47b96f8365c8e38f4): implementation SHA256 `9ee5f005fa278268015b3ec619f71a2f524b181a64f3f4b79b44fd67db5cd136`.
- Original [7b63df7](https://github.com/pingdotgg/t3code/commit/7b63df7c8123148b2110163fa73c72aa891de85a): reference identity only; not the tested implementation.
- Both variants retain exact head test content, SHA256 `9f0f42b2f87db068f619281fa190aae53e4395e095c5229031a95cf27c897477`.
- Final driver SHA256 `b18151f95425ae7ac7149d3c6168608e1c60091a70cf8ba8fea393d6032e6676`.
- Node v26.12.1, Effect and @effect/vitest 4.0.5, Vite+ 1.3.0, Vitest 5.0.5, Linux x86_64. Installed worktree dependencies reused through symlinks; no install needed. Exact runtime/OS, commands, start/end UTC, exits and process IDs are in `environment-cleanup.json` and `commands.json` in the archive.

Final captures ran 2026-10-23 06:27:06–06:27:12 UTC. Source hashes match the matrix. SHA256SUMS covers every archived source, test, config, script and transcript. Source bytes are exact git blobs; no import rewriting was needed because staged node_modules resolves the workspace shared package.

## Exact test results

Same head tests and dependencies, isolated minimal test config (no repository setup hooks), against each implementation:

- Red focused: exit 1; 2 failed, 6 skipped. Both failures are `signal?.aborted` undefined versus expected true.
- Red full file: exit 1; exactly 2 failed, 6 passed.
- Green focused: exit 0; 2 passed, 6 skipped.
- Green full file: exit 0; 8 passed.

“Full” means the entire affected eight-test file, not the repository-wide suite. All exact stdout/stderr is retained; no excerpt substitutes for a transcript.

## Runtime coverage and observed outcomes

The driver imports exported `mcpToolSession`, runs `.call` after initialize and initialized notification inside `Effect.scoped`, and imports exported `runAcpMcpStdioBridge` with PassThrough input. Stdio sends initialize, initialized and tools/list in sequence through its real barrier. Both verify session/protocol replay. Each target request is correlated by path, row, JSON-RPC id/method and socket port. Native fetch is wrapped only to log signal/header/rejection/body-read milestones; the same rejection object is rethrown. Synthetic authorization and harmless `{ok:true}` results only.

- T1, both exports: request consumed, headersSent=false, then owning fiber interrupted. Red stays open with no abort event throughout 150 ms, followed by explicit response release. Green records one target signal abort, native fetch AbortError, target response/socket close with headersSent=false, released=false and cleanup=false. Session DELETE is logged separately and acknowledged immediately with 204.
- T2, both exports: real JSON headers and partial body sent; native fetch resolution and actual response.text invocation recorded before interruption. At both revisions interruption completes within the deadline, no target abort occurs, and target stays open during 150 ms observation. Remaining body is explicitly released afterward. This confirms unchanged JSON body lifetime, not SSE cancellation.
- T3, both exports: complete tool result / target stdio output observed first; session scoped finalizer completes normally, stdio EOF and Fiber.await complete, then interrupt completed fiber. Result/output retained, zero aborts/rejections, no duplicate target output.
- T4, both exports: two interruption requests against the same pending owning fiber both finish within the deadline. Green observes one abort event, one target socket close before headers/release/cleanup; red stays open 150 ms until release. No duplicate target stdio output or unhandled rejection. One abort event is not a direct count of abort() calls; installed Effect source trace records the per-async-operation controller and interruption finalizer, which runs with interruption disabled.
- T5, both exports: no interrupt during operation; expected result/output, normal handshake replay, EOF/completion, session DELETE; zero aborts/rejections/unhandled rejections.

All 20 cases and both driver runs exit successfully. No unhandledRejection during the full captured lifecycle. Requests, gates, interrupt/join and fixture stop are bounded to 5 seconds; external diagnostic observations use 150 ms windows. No sleeps/polling were added to repository tests.

Request-class collapse follows the auditor matrix: initialize and notification share the changed session send site with tools/call; the existing focused test blocks initialize and the runtime reaches tools/call explicitly. Stdio tools/list exercises its changed forward site; initialize/initialized replay is captured in every case. Existing passing tests cover silent notifications, parse input, SSE forwarding, concurrent client replies and protocol cancellation forwarding. Those do not turn transport interruption into an MCP notifications/cancelled claim. DELETE is a separate unchanged transport. No extra Cartesian suite is claimed.

## Ownership, cleanup and boundaries

Runtime PIDs: red 1030917, green 1031430. Each of 20 fixture listeners records `ss -H -ltnp`, matching PID and `/proc/<pid>/cwd` before use; finalizers check cwd before closing their owned listener and tracked connections. Final verification confirms both processes exited and all 20 ports have no listener. No kill commands were needed. Target pre-header closure is recorded before cleanup; completed keep-alive sockets may close during cleanup, as explicitly labeled. Final free disk 8,921,518,080 bytes, above the 5 GB stop threshold; capture archive is small.

No home directory was needed for direct exported effects. Capture subprocesses set `T3CODE_SEED_DEV_DATA=0`. No real database was read, copied or opened; no T3 app/server was launched.

Excluded claims remain: CLI/OS-signal cancellation source, every provider, remote tool rollback, protocol cancellation, JSON/SSE body abort. The outer CLI runPromise has no in-process interruption hook; direct-effect evidence does not establish one. No SSE body timing capture is claimed. Capture acceptance is complete within these boundaries; separate auditor inspection remains the auditor's work.

## Auditor rerun

The existing driver and staged source remain usable. From the authorized worktree:

```sh
cd /home/ubuntu/wt/pr-14419
sha256sum /tmp/ev-14419/capture/runtime.mjs
node /tmp/ev-14419/capture/runtime.mjs red T1 session
node /tmp/ev-14419/capture/runtime.mjs green T1 session
```

Use `stdio` instead of `session`, or T2/T3/T4/T5 instead of T1. Omit row/path to run all ten cases for one implementation. Every runtime invocation logs implementation revision/content hash, driver hash and ownership.

For fresh staging after unpacking the archive into `/tmp/ev-14419/capture` (same existing pinned worktree/dependencies):

```sh
python3 /tmp/ev-14419/capture/prepare.py
vp test run --config /tmp/ev-14419/capture/red/vite.config.ts --root /tmp/ev-14419/capture/red --reporter verbose
vp test run --config /tmp/ev-14419/capture/green/vite.config.ts --root /tmp/ev-14419/capture/green --reporter verbose
```

Append `-t 'aborts the pending fetch'` for focused runs. Red exit 1 is expected. `run-captures.py` reproduces all six exact commands but overwrites local transcripts, so preserve the published originals first. `finalize.py` verifies all 20 case/cleanup records, process exit, port closure and clean product status, and regenerates metadata/hash inventory.

## Final-head refresh provenance

The supplied final head merges main at the pinned base. Server implementation and head-test hashes are unchanged from the previous capture. All six capture commands and all 20 cases were nevertheless rerun from exact git-blob staging. Previous local artifacts are preserved under `/tmp/ev-14419/prior/pre-final-refresh`; old public uploads remain intact. Parent `/tmp/ev-14419/final-*.txt` logs were neither overwritten nor included in this archive.

## Fresh ghdrop uploads

[red-tests-focused.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a9c2da06de60555a/red-tests-focused.txt)
[red-tests-full.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/167e21b1c8169fea/red-tests-full.txt)
[green-tests-focused.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ce655d508cde8d89/green-tests-focused.txt)
[green-tests-full.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/adad6e32bb57ee7f/green-tests-full.txt)
[red-runtime.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/23ecb8290faa73fc/red-runtime.txt)
[green-runtime.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/519d3b3c753edc7d/green-runtime.txt)
[runtime.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2add569f47ff22ed/runtime.mjs)
[SHA256SUMS](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d6433e4ce3d8a9e2/SHA256SUMS)
[capture-files.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3e8698b2dd9baea2/capture-files.tar.gz)
[matrix.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e4aa0353460ff0c8/matrix.md)

## FRICTION

- No capture failures or driver behavior changes. Only revision pins changed; exact final sources and tests were staged outside the product worktree.
- Minimal test config avoids repository setup and database hooks. Full means all eight affected-file tests.
- Parent final-*.txt logs are outside this refresh and untouched. No merge or independent audit verdict is claimed.
- No code changes, pushes, PR writes or watches. T3 thread PR registration is metadata only. All synthetic HTTP fixtures stopped; no real database access or seeding.