# Issue 21578 runtime evidence

Base: `8d3b061121fe65669d5d43d011b48cd3027a4108`. Head: `521eeb4349f8795bdbede9a5af73d1e79556f1fd`.

Server-only evidence matrix: empty. Lead proof audit pass 1 found no render paths. No captures, recordings, or seeding commit required.

## Runtime observations

Existing lead-owned isolated server and browser were left untouched. Lead reports successful real-browser Codex GPT-6-Astra turns on both revisions. Conversation content is excluded.

| Revision | Recovery spans | Min (ms) | Max (ms) | Mean (ms) | Median (ms) |
|---|---:|---:|---:|---:|---:|
| before | 8 | 45.116551 | 90.649134 | 64.417225 | 63.282523 |
| after | 8 | 18.854866 | 32.158883 | 27.152363 | 28.290088 |

The JSON transcript contains exact span IDs, trace IDs, Unix nanosecond start times, UTC start times, and durations. Export uses an explicit metadata allowlist; no attributes, events, links, exit objects, payloads, conversation text, or credentials are exported.

One baseline-ID candidate (`0dfd0df7e017fa96`) belongs to a different trace and is absent from the baseline summary. It is documented as excluded rather than counted among the eight baseline spans.

## Read-only database confirmation

SQLite was opened with `mode=ro`. Migration 68 is `UnfinishedTurnItemsIndex`; its status-only partial index exists on `thread_id` for statuses `pending`, `running`, and `waiting`. Exactly 17,517 fixture rows match the fixture prefix; all are completed command executions in one thread. No row payload was selected.

## Reported tests

Implementer: green 64 tests across 19 files. Independent reviewers: green 39 tests across 2 files. These counts are lead-supplied; this evidence step did not rerun tests and received no original test output. Exact red assertion remains pending the implement report; it cannot be truthfully reconstructed from the supplied runtime files.

## Reproduction of this export

Retain the original three lead input JSON files and trace snapshot. Run `python3 export.py` in the same local environment; the script verifies all selected span IDs, the baseline duration sequence, after metadata, and read-only database counts before exporting. The JSON records the consumed trace snapshot SHA-256 and byte count; an actively appended trace will have a different hash. This script reproduces evidence extraction, not the original app interaction.

## Limitations

- Synthetic completed projection rows are performance fixtures, not event-replay business proof.
- Same isolated thread reused; after includes the previous successful user/assistant turn.
- Single before/after interaction with eight internal recovery spans each, not eight independent trials.
- Timings are indicative and do not establish exact reporter latency or end-to-end speedup.
- Commit attribution and interaction outcome are lead-supplied; trace metadata alone does not encode commit identity.

## FRICTION

The baseline ID file contains nine candidates but only eight belong to its measured interaction; reconciled using trace identity and baseline durations. The exact red assertion and raw test reports were not supplied. No source edits, commits, browser operations, server launches, or server stops were performed.