# Issue 22577 — updated server-only evidence index

This supplement supersedes the pending/raw-output and correlation limitations in the [previous index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/236c12ba978a0529/22577-evidence-index.md). Evidence captured 2026-10-21 UTC. No source edits, browsers, servers, pushes or forge comments performed by this capture agent.

## Exact revision applicability

- Base: [70e5e384da](https://git.cafe/pingdotgg/t3code/commit/70e5e384dab96a1d1e4a253ee5a9415ed5d8ce30).
- Current head and fresh focused green: [55ac62ba33](https://git.cafe/pingdotgg/t3code/commit/55ac62ba338e291db024ce7a7b56d39603171d5e).
- Actual after-runtime proof: [d1736d5221](https://git.cafe/pingdotgg/t3code/commit/d1736d522188655d5dd4215ffbff4d9d17065c63). This is not a runtime run at current head. The only changed path from this revision to current head is apps/server/src/orchestration-v2/runtimeLayer.test.ts. Orchestrator.ts is byte-identical (Git blob 9f7973ab7868c4dd91c0b6f4caa96714f9bc0f9f at both revisions); applicability to current implementation follows that identity.
- Supplied tester red capture uses detached [0df2b9c89d](https://git.cafe/pingdotgg/t3code/commit/0df2b9c89d9195a78c669c29351c85de4e33e58a), whose Orchestrator.ts is byte-identical to base (Git blob 8c8620088fa9431c3360552688a3df2a4f20557b). New-case red explicitly copies current-head tests onto that base implementation. Red was supplied by tester, not rerun by this capture agent.

## Visual/redraw scope

Visual matrix remains empty: 0/0 affected rows. The complete base-to-head scope is server orchestration policy plus deterministic runtime-layer tests. No changed render path, style, layout, client interaction or redraw behavior requires a visual fixture. This preserves the server-only exemption in the [independent audit](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/159e126c06e05939/22577-proof-audit-pass1.md). No screenshots, recordings, visual diff or redraw-performance claim is made. No browser was opened for this supplement.

## Genuine red and fresh green

[Original raw red](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e63fc15caae7900b/22577-raw-red.txt): original named selection regression, exit 1, expected gpt-5.5 but received gpt-5.4 after queued completion starts.

[Four-case raw red](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/43396365261fece7/22577-raw-red-new-cases.txt): current-head tests with base implementation, four failures / 83 skipped. Model and instance cases retain old selection; usage-limit resume incorrectly remains blocked by old provider; disabled-current-instance case starts instead of staying queued. These are genuine tester output artifacts, replacing the previous report-only attribution.

[Fresh raw green at current head](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd01432aaecea113/22577-green-focused-head.txt): command from apps/server: `../../node_modules/.bin/vp test run src/orchestration-v2/runtimeLayer.test.ts -t 'delegated completion' --reporter=verbose`. Exit 0, four passed / 83 skipped (87 total), duration 5.88s, start 23:32:55 UTC. Artifact includes full revision, cwd, command, individual test names and raw output. No source changes made.

R1/R2/R3 coverage:

- R1: queued delegated completion follows the current model and preserves the thread selection.
- R2: current instance routing follows the selected instance; run, attempt, active provider thread, root node and checkpoint scope agree. Same-instance model change continues the parent's provider thread; alternate instance uses its own provider thread. This is deterministic cross-instance proof, not a live cross-provider adapter claim.
- R3: a limited current provider holds the completion and rejects queue.resume; switching to an available instance allows resume despite the old provider's limit. A project-disabled current instance keeps the completion queued and held.

The tester additionally reports all 87 tests passed, root lint clean and typecheck passed. Those broader results were not rerun or independently captured in this supplement; the linked fresh artifact independently proves the four focused cases only.

## Runtime causality and actual native model

[Before runtime](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/df63f6dae4a4ced8/22577-reproduction.txt): qualifying attempt 2 queues completion ordinal 3 at 22:48:35.387Z, persists Sonnet at 22:50:09.131Z, then starts completion on Opus and rolls selection back at 22:50:19.095Z. Attempt 1 is not a queued-completion reproduction.

[After runtime summary](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ad23a04948d691c7/22577-after-runtime.txt) and [correlated after-runtime supplement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/958307d955259afc/22577-after-correlated.txt) apply to d1736d5221. The correlated artifact supplies the previously omitted message/run IDs, parent run 4, generation 1, delegated task ID, provider-thread/session/native-thread identity and attempt identity.

Ordered milestones: parent 4 completes 23:19:30.927Z; blocking user run 5 starts 23:20:01.659Z; completion run 6 queues on Opus at 23:22:03.967Z with the correlated delegatedCompletion message; thread persists Sonnet at 23:22:27.473Z; run 5 completes 23:24:06.409Z; the same run 6 starts on Sonnet 23:24:06.816Z, runs 23:24:07.523Z and completes 23:24:12.046Z. Final persisted selection remains Sonnet in the summary; no rollback appears in the supplied sequence.

The correlated fixture contains one distinct completion message, one distinct completion run and one attempt. Running/completed attempt binds native session dbdd699e-a2ff-4960-85cb-2f75b81fe259. Native response metadata at 23:24:11.989Z explicitly reports claude-sonnet-5-5 in that session inside completion run 6's attempt interval. The earlier native Opus response at 23:24:06.352Z belongs to blocking run 5. Session projection's startup Opus label is not treated as the executed model. This closes the supplied proof's runtime-correlation and actual-native-model ambiguity for this fixture; it does not establish universal exactly-once behavior.

Reproduction setup: complete an Opus parent with an asynchronous child still outstanding; run a second blocking user turn; let child completion queue; persist Sonnet with threadEnvironment.updateMetadata (picker draft alone is insufficient); let blocking turn finish and inspect the correlated completion. Existing deterministic tests seed their own state; no separate UI seeding artifact is needed.

## Limits and friction

No live runtime rerun at 55ac62ba33. No native desktop/mobile, remote/relay/tunnel, or live cross-provider runtime exercise. Existing same-provider native runtime plus focused deterministic instance/guard coverage is the supplied proof. No new visual or redraw claim.

FRICTION: vp is absent from ambient PATH; used the existing worktree node_modules/.bin/vp, without installing anything. Raw-red and runtime-correlation gaps from the prior index are now addressed by public artifacts. Worktree remained clean. No other friction.