Proof audit pass 1
Issue: [#23050](https://git.cafe/pingdotgg/t3code/issues/23050)
Workspace: /home/ubuntu/wt/issue-23050
Reviewed head: [111dadc991](https://git.cafe/pingdotgg/t3code/commit/111dadc991e38448d78cc76e8d13e02ed1e0abf5)
Base: [d27c6067a9](https://git.cafe/pingdotgg/t3code/commit/d27c6067a9162150f654e38f455a7a936ef79178)
Branch: fix/issue-23050; clean tracked worktree at inspection.

Scope and UI fixture matrix
UI matrix: EMPTY (0 rows).
The complete diff changes only packages/provider-pi/src/server/adapter.ts and its existing adapter.test.ts. No client, UI entry point, wire contract, navigation, style, or timing presentation changes. Web, desktop, and mobile consume the existing fork operation; their layouts and interactions are unchanged. No screenshot, interaction recording, timing recording, browser, or server launch is required for this proof matrix. Other provider adapters and connection transports are unchanged.

Non-UI proof fixture matrix
P1 — Pi native fork at completed turn 1 while the next turn is running, has startedAt, has completedAt=null, and still has a synthetic/uncaptured session-tree reference. The active session tree contains u2 and later steering input. Expected: fork selects u2 rather than steering input, returns the forked native session, retains precisely turn 1 ancestors, and excludes turn 2.
P2 — Existing completed uncaptured boundary still uses its finite end: a later prompt cannot substitute for a canceled turn that ended before that prompt landed. Existing adjacent regression covers rejection, followed by successful finite-window resolution. This is retained coverage, not a request for another new test.

Implementation/proof assessment
The implementation removes only the completedAt-null rejection, retains the startedAt-null guard, and uses Infinity as the upper timestamp bound for the current session-tree snapshot when completedAt is null. Completed boundaries retain their finite upper bound. The branch walk and earliest qualifying user-entry selection remain unchanged.

The single new regression named "natively forks before a running turn whose start entry was not captured yet" exercises P1 through runtime.forkThread with the fake Pi protocol. It asserts successful fork, native session identity, fork entry u2, and retained turn-1 ancestry. It uses no sleep, polling, or wall-clock timeout. [c208208bad](https://git.cafe/pingdotgg/t3code/commit/c208208bad47c73f3208242ca07635108e8a2a5b) contains that test without the implementation fix; its adapter implementation matches the base. The test is unchanged between that red-test commit and the reviewed head.

Existing /tmp/23050-before.log is a failing temporary runtime-test transcript using the fake Pi adapter. Exact protocol detail:
"Pi turn 2 has no captured session-tree entry and none was found in the session tree"
It records ProviderAdapterForkThreadError wrapping ProviderAdapterProtocolError, then expected Success versus actual Failure. This establishes the reported pre-fix failure through the adapter runtime; it is not a transcript from a live Pi executable. Its temporary test name differs from the committed regression, so it does not independently establish an execution of the exact committed test at the red-test SHA. Red-first commit provenance is verified structurally; the supplied assertion of the named test's red run is not independently rerun here.

Requirement status: NOT YET FULFILLED at pass 1. The red regression and before transcript match the failure, but the requested after runtime transcript is explicitly pending. No green runtime result at the reviewed head was supplied or produced in this read-only proof pass. /tmp/23050-baseline.log predates the fix and shows three existing tests passing; it is not after evidence for the new regression.

Outstanding proof row
P1-after — Lead must supply the short after transcript tied to the reviewed head, showing the same running/uncaptured boundary succeeds and selects u2 while retaining turn 1. Include the focused committed regression's green result at that head. If claiming an independently executed red result for the exact committed regression, preserve that result at the red-test SHA as well. No UI evidence is needed.

FRICTION
- No server, browser, forge command/public write, or repo-wide check was performed.
- No worktree edits, new fixtures, evidence captures, or test runs were performed. Only this requested report was written outside the worktree.
- The available before transcript is fake-protocol runtime evidence, not live-provider integration evidence; do not label it as the latter.
- This is pass 1 only. Pending after evidence must be assessed in a later pass; no pass-2 completion or merge-readiness claim is made.

VERDICT: INSUFFICIENT P1-after