# Evidence index: Pi compaction submission

Issue: [#24227](https://git.cafe/pingdotgg/t3code/issues/24227).
Base: [27e0c684c5](https://git.cafe/pingdotgg/t3code/commit/27e0c684c585f955252d65f06206dd11a143bff5).
Head: [588b1c29f0](https://git.cafe/pingdotgg/t3code/commit/588b1c29f099b69924824c2304b6ba66dbb838e7).
Red regression commit: [7737dcc721](https://git.cafe/pingdotgg/t3code/commit/7737dcc72102237a5665a82c33a55c7a87cabe64).

## Required proof and provenance

The lead reports the regression test failed against base with zero delivered prompts versus one required, and 92 tests passed at head. These are lead-reported execution results, not rerun by this evidence-capture agent. Raw red/green execution logs were not among the supplied artifacts; the test commit is source evidence, not an execution transcript. This limits independent verification of the test half of the proof requirement.

The lead performed the integrated real web/Pi 1.4.0 replay. Browser: Chromium 154, viewport 1280x800, no emulation. This agent inspected and uploaded the artifacts; it did not start a browser or server or independently repeat the runtime replay.

- [Before runtime: failed turn and raw Pi rejection](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5222e41c045983dc/before-runtime.txt).
- [After runtime: ordered compaction, submission, native receipt, assistant reply and completed turn](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/206235b730da78a4/after-runtime.txt).
- [Exact fixture and reproduction steps](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0220db8becf6bbe8/reproduction.md).
- [Native Pi extension driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/464c2348b4321eff/compact.ts).
- [Original proof audit / empty render matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/16a827664c136d77/proof-matrix.md). Historical audit: its runtime-PENDING verdict predates the supplied after transcript; its statement that the before transcript lacks raw rejection predates the expanded before transcript uploaded above.

The after transcript records session_before_compact at 10:27:58.417460 UTC, app submission at 10:28:02.225, persisted compaction at 10:28:31.555, native user receipt at 10:28:31.565, assistant RECEIVED at 10:28:33.831, native marker count 1, and completed provider turn at 10:28:33.845. This supports single-message delivery after compaction in this replay. It does not demonstrate multi-message FIFO runtime behavior.

## Render proof applicability

Matrix 0/0: server-only adapter and adapter tests; no changed render paths. No captures, pixel diffs or recordings are required by the server-only proof rule. Screenshots below are optional illustrations, not a pixel-unchanged claim. They use different fresh threads and sidebar state.

Before — illustrative rejection:

![Before: Pi compaction rejection and failed turn](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a331ab48ed786a69/browser-screenshot-localhost-mvm8elua-8a7dea4c.png)

After — illustrative successful reply:

![After: compacted context and RECEIVED reply](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e739a65881917d6c/browser-screenshot-localhost-mvm9am4r-fef7c55b.png)

## Limits and handling

Only the supplied web/Pi detached-compaction replay is runtime evidence. Desktop, mobile, remote/relay/tunnel modes, other providers, FIFO ordering, stale-owner dropping, and successful post-Stop restart were not independently runtime-verified here. Other adapter branches have source/test evidence to the extent described in the original audit; do not generalize this single-marker replay to them.

Text artifacts and both full screenshots were inspected before public upload. No apparent credentials, private customer content or pairing links were found. The screenshots contain fixture text and local development UI. No repository edits, commits, forge writes, browser/server starts or server stops were performed.

FRICTION test-provenance|Raw red/92-green logs were absent from the supplied evidence|Save and attach test command/output alongside runtime evidence; distinguish reported results from independently inspected logs.
FRICTION visual-comparability|Before and after use different fresh threads/sidebar state|Label screenshots illustrative and keep the render matrix 0/0; make no pixel-diff claim.
FRICTION coverage-claims|A single native marker proves one delivery, not multi-message FIFO or other clients|Scope runtime conclusions to the observed replay and distinguish source/test coverage.