# Proof audit pass 2 VERDICT: INSUFFICIENT. Required evidence remains a pending gate for the dispatcher. No merging. Reviewed clean worktree /home/ubuntu/wt/issue-24659 at 27039417731e85cc54026b1b960df4e9a08e0ad4 versus c79ef8c454. Read the entire two-file diff, pass1 matrix, evidence index, uploaded drivers, wire logs, manifest, DB proof script/results, all ten recordings through extracted temporal frames and terminal frames, and all still/diff sets. No source edits, commits, pushes, comments, merges, watches, or server launches/stops. Temporary audit artifacts are under /tmp/proof-pass2. Reused existing server 5778 and isolated .t3; own preview tab and fresh pairing. ## Confirmed evidence defects / required retakes - Row 2 baseline provenance is insufficient for an exact-base claim. The retained 6.47s baseline has one logged blocked send and no durable message, but no served-module identity. The index explicitly reports later base-on-disk captures serving head code. Branch checkout/diff labels cannot establish which client module ran. Retake base in an independently verified base-serving session, recording the loaded normal module's identity and the reconnect/no-resend result. Do not substitute the excluded later baseline clips. - Row 2 original desktop clip ends at 10.14s with OK appearing but Working/Stop still present; it does not establish the final settled controls. The 43.39s narrow-dark clip ends in Thinking/Stop. Retake narrow-dark through stable completion. My independent desktop head recapture supplies a settled head sequence, but does not repair baseline provenance or the narrow clip. - Row 1 proves one image and its image-context record, not a separate file/context reference. The retry cases all carry empty attachments. Thus attachment/context preservation across retry and no reupload are unverified. Complete the combined attachment/context + reconnect case, or explicitly narrow the claim; do not call the original row fully complete. - All ten uploaded MP4s have video streams only, no audio narration; they also lack explicit scenario/SHA/command-ID captions. The index maps filenames to scenarios, but the clips alone are not narrated exact-SHA proof. Add a timestamped narration/caption sidecar or annotated exports identifying fault, send, reconnect, and final state. No need to pretend silent clips contain narration. ## Row accounting 1 PARTIAL: healthy button/image send valid; separate context and attachment-bearing retry missing. 2 PARTIAL: head resend behavior valid; independent desktop recapture succeeds. Exact-base provenance and completed 390-dark sequence require retake. 3 CORE VALID: two identical payloads on sockets 0/1; accepted_at 06:11:31.447Z, first send 06:11:31.430Z, retry 06:11:37.009Z. First commit therefore precedes retry, independently confirming committed-loss rather than merely pre-delivery loss. One accepted receipt/message/completed run/provider attempt; recording settles to OK. Suppression stops all frames, so the separate projection-before-receipt ordering condition and every internal outbox effect are not proved. 4 CORE VALID: three identical payloads on sockets 1/2/3; one durable completed run and provider attempt; clip shows recovery and final OK/idle composer. 5 CORE VALID: three identical attempts, restored draft, optimistic row removed, eventual connected idle composer, zero durable messages/runs; no fourth in supplied observation. No persistent post-recovery error banner is shown. Do not call this proof of a persistent error or unlimited future observation. 6 CORE VALID: actual typed OrchestrationV2DispatchCommandError/ProjectionStoreThreadNotFoundError response, one dispatch, error toast and restored draft, zero durable messages/runs. This proves the supplied business-rejection case, not all possible error classes. 7 PARTIAL: Settings navigation plus resend is valid. No acknowledged cancellation or fiber-interruption proof; cancellation semantics remain unresolved. 8 UNVERIFIED: steer/restart and modern/legacy mode coverage. 9 PARTIAL: queue_after_active identical retry and one durable completed run valid; app action, queue cancellation/reconnect, and full settled queue lifecycle missing. Clip ends with queued item already released into a running turn, not final idle. 10 UNVERIFIED: Resume/manual continuation/held queue. 11 UNVERIFIED: lifecycle compaction and compact-before-send. 12 UNVERIFIED: same/new-thread plan implementation. 13 UNVERIFIED: native iOS/outbox. 14 UNVERIFIED: native Android/outbox/cancellation. 15 UNVERIFIED: Electron shell. 16 UNVERIFIED: remote/relay/tunnel. Rows 8, 10-16 may be named as unverified limits in a narrowed PR claim, but cannot be counted complete. Rows 7/9 must remain partial. Any requirement retained from the pass1 matrix remains pending until proved or explicitly removed by the dispatcher. ## Visual and durable checks No confirmed stale overlay, ghost, missing asset, or unrendered final still was found. Transient upload/loading/reconnect states in recordings are real intermediate states. The top timeline fade is not evidence of a ghost. The fixture is intentionally a tiny blue image, and renders as blue. Recomputed pixel differences: original 1440-light pair 676 changed pixels / 167 above channel threshold16, bbox x226..227 y0..678; repeated pair zero. 1440-dark, 390-light, and 390-dark repeated pairs zero. These prove within-state still stability only. Before/after histories, scroll positions, timestamps, branch labels, and checkpoint summaries differ; no before/after CSS equality claim is valid. Desktop still PNGs are 1280x800 although manifest reports CSS1440x900; desktop videos are 2880x1800. Narrow PNG/video are 780x1688, consistent with CSS390x844 at DPR2. Correct physical labels or retake exact-size stills if required. Four theme/width stills are present; behavioral runs do not cover all four combinations. All uploaded retry payloads compare equal, including command/message IDs. Requeried the isolated DB read-only: all ten uploaded DB message/run results still match. The uploaded script counts provider attempts under the misleading key attemptEffects; it does not count every internal effect. Preserve that limitation. Public drivers are usable traces and no committed fixture/driver is needed, but the files alone omit the complete setup/action sequence and exact loaded source identity. ## Independent head rebuild Own tab tab_3_73aaaf7c-35e8-4617-8bdc-88f24380768b; CSS1440x900 light; uploaded driver.js installed unchanged, __mode=pre, __drops=1; Enter send into existing stopped evidence thread. Unique harmless prompt ends 'pass2 independent pre-delivery proof'. First attempt 06:22:58.332Z, second 06:23:01.289Z; identical payload. Old socket is logged -1 because it predates wrapper installation; replacement is 0. First attempt is explicitly dropped before native send. Accepted receipt 06:23:01.320Z; one message, one run completed 06:23:07.332Z, one provider attempt. Captured continuous recording through settled OK and idle Send control. commandId a72aa86e-b709-4fe0-9af9-7dc6518d92ba; messageId dfdaf46f-4e6a-4d68-86cf-34d2eabe9263. Trace: /tmp/proof-pass2/independent-trace.json. Final frame: /tmp/proof-pass2/independent-final.png. Recording: /home/ubuntu/.t3/userdata/attachments/thread-delegated-task-command-3amcp-3a3235bc45-87b2-424c-bab7-d2dbdf4a574e-3adel-cac6fa37-f726-407e-81b4-beba471a99f9-mp4.mp4. Fetched served @fs commands module contains retry, SHA256 d0c24dae8e1f7f51478a808c4b2acad0a044d8901a180a3f0942b10537f14899. This supplemental fetch is not a historical loaded-module attestation for earlier captures. ## FRICTION Fresh pairing needed despite shared Default profile; successfully recovered using pair --base-dir /.t3. Initial proxy502/503 recovered without server intervention. Python urllib download was403; curl succeeded. Preview resize requires mode=freeform; incorrect initial modes rejected harmlessly. Contact sheets scale narrow frames too small: original frames do include composer, correcting my initial commentary; the real defect is early termination before settling. No source mutation or base checkout performed. No tests executed in this evidence-only audit; focused test gaps from pass1 remain lead-owned. Native clients, Electron, remote endpoint, resume/plan/compaction fixtures absent from supplied proof. No broader surface success inferred.