# Fresh capture proof audit, pass 2 — INSUFFICIENT

Scope: [#17031](https://git.cafe/pingdotgg/t3code/pulls/17031). Reviewed after [34fcb08](https://git.cafe/pingdotgg/t3code/commit/34fcb08ce3146c0b6424d676b1d05ec8b49014c4), before [977a188](https://git.cafe/pingdotgg/t3code/commit/977a18823cf8e067476193e29947e2d3950a25f2). This is a capture/provenance audit, not a code audit or merge approval. Source HEAD was independently verified at the exact after SHA with empty git status. No source changes, browser/server/provider launches, delegates, tests, or forge mutations performed by this auditor. Only isolated /tmp runtime data was read; no live T3 home access.

## Independently checked evidence

- Read fixture matrix and evidence manifest; inspected local artifact inventory, all eight named R1 PNG dimensions, and visual contact sheet. Exact images are 1440×900 and 390×844 for both themes/revisions. The additional initial-observation image is not a ninth matrix cell. Before has completed reply and no page; after shows readable OPENCode-EXECUTE page and final reply, in both themes, with whole app viewport. No obvious unsettled placeholder in the eight matrix images. A static capture cannot independently prove font readiness or consecutive stable heights; those remain capture-agent observations.
- Real OpenCode before and corrected after row exports exactly match canonical persisted turn-item payloads read through SQLite mode=ro. Both are completed execute calls invoking discovered t3-code html_render; their string outputs contain htmlRender reference/title/heights/message. HTML extracted from both actual call inputs has identical SHA256 a62e076d2984ccdb86586b961fc589bcdf9fd8686ccb94689c5537518f8dea76. Provider protocol logs exist for both thread IDs and record OpenCode sessions. This corroborates real provider provenance, rather than a screenshot-only or replay claim. Capture-time revision binding, provider version and process environment are historical manifest claims, not derivable from pixels alone.
- Unsupported first after call is genuinely double-wrapped and correctly excluded. Correction/history differences are disclosed; do not present this as identical prompts or proof of insertion without reload.
- ffprobe confirms playable H264-sized 1440×1000 recordings of 30 seconds after-flow, 25 seconds before-flow, and 140 seconds continuous after. Extracted and viewed frames from after-flow at 2/9/18/28 seconds, before-flow at 2/20 seconds, continuous after at 20/100 seconds. Observed desktop Count:2, narrow Count:1, before absence, full-screen Count:0, and a transient T3 loading screen inside the recording. ffmpeg log identifies x11grab. These are moving desktop capture artifacts, not settled PNG substitutes. Sampling corroborates interactions but does not independently certify every transition or 0→1 sequence. Before scripted reload locator failure is explicitly disclosed; do not claim its assertion passed. Loading during intentional reload does not invalidate separate settled PNGs.
- SQL harness actually imports ProjectionStoreV2 and migrated SQLite layer, inserts synthetic completed execute rows, calls getThreadAttachmentIds. Saved before result selects neither own wrapped nor bare; after selects both and excludes foreign/isError fixtures. This is genuine SQL-service integration with synthetic input, not a live provider or guest test. The output is observational (no failing assertions/exit checks in harness itself); test success cannot be inferred merely from process exit.
- Lifecycle JSON labels manually inserted bare/foreign/isError rows APP REPLAY, wrapped row REAL PROVIDER. Independent canonical SQLite query confirms owner deletedAt 2026-10-27T01:32:47.816Z and foreign thread undeleted. Filesystem check finds surviving foreign HTML and no corrected own HTML. Historical before-existence and actual UI deletion action are retained transcript claims, corroborated by final state, not independently replayed here. Failed fixture deliberately survives; this does not establish every failure-status case.
- Test transcript totals agree with manifest: after 192 focused + 41 mobile + 78 existing server; before 223 across six paths. Files contain aggregate summaries, without individual assertion names, command echo, SHA or exit status. Exact commands/revisions/status are supplied by manifest, not self-authenticating logs. Before lacks head-only cases and must never be described as passing them.

## Remaining minimum rows and retry priorities

1. R2 mandatory REAL ACP: no successful ACP tool invocation exists. Grok 1.0.67 advertises grok.com while T3 requests xai.api_key; generic registration is unauthenticated. This is a concrete provider/auth compatibility blocker, not proof the HTML patch is broken. Retry at least one actual ACP agent on both exact revisions, preserve native updates and persisted output, add eight whole-viewport PNGs and before/after interaction recordings. The existing locally configured OpenCode generic ACP route is a sensible retry candidate: running OpenCode through ACP can satisfy the minimum actual-ACP criterion if it really traverses ACP initialize/session/tool_call/update and emits html_render. SDK OpenCode R1 cannot substitute. Label it OpenCode ACP, never Grok; Grok-specific identity/OkayOutput and live nested Grok remain unverified. This route was not executed or independently established as authenticated by this auditor. If the literal Grok requirement is retained, alternate ACP only satisfies the general ACP minimum and Grok stays open.
2. V1 guest: actual guest endpoint service and actual guest UI still absent. Connect requirement is infrastructure/configuration evidence, not a passing guest row or an HTML code defect. Retry with an authorized working Connect setup or another supported real guest route; obtain eight settled before/after width/theme PNGs plus recordings. Blocker screenshots do not count.
3. S1: add actual guest serving (own allowed, foreign denied), unchanged MCP App/unrelated-upload control, and before-revision filesystem deletion transcript. Keep synthetic SQL/application insertions labeled replay. Current after deletion and real SQL reads are useful partial proof, not the full lifecycle matrix.
4. M1 fallback: recorded emulator -list-avds exit127 and both-revision mobile logic tests support explicit native-unverified fallback. An absent executable is not an attempted emulator boot. Native WebView/theme/interaction/reopen remain unverified. If matrix requires an actual boot failure rather than unavailable tooling, obtain exact available-host boot attempt or explicitly accept unavailable-tooling fallback. Narrow web never proves native.
5. Tighten publication proof: attach command/SHA/exit provenance for summarized tests and frame/interaction index if claiming complete recording acceptance. No need to rerun already supported R1 screenshot cells solely because ACP/guest are missing.

## Claims and limits

Verdict INSUFFICIENT because mandatory evidence remains missing: only 8 of 24 required settled success PNGs exist; 8 ACP and 8 guest are missing. No demonstrated code bug is established by this capture audit. Do not infer ACP, guest, native, desktop shell, relay/tunnel, extra providers, live nested child, or full S1 green from unit tests. Do not call the before reload script successful, the double wrapper supported, or the recordings uniformly settled. No false settled-state claim was found for the eight named R1 PNGs at inspected resolution. Execute-output trust is accepted maintainer policy and is not a finding.

No new PR source tree was reviewed after the exact SHA above; parent publication preserving that tree through merges must preserve this scope explicitly.

Auditor: gpt-6.1-sol through Codex.