# Proof audit, round 2

PR: [#23813](https://git.cafe/pingdotgg/t3code/pulls/23813).
Head: de44c6e1b7bee2bb7d41e124bdf8c22b8911075c.
Merge-base: 22840ab943b8418744b48223411e46ba4fdcef1a.
Machine: fs-36. Fresh single installed worktree: /home/ubuntu/wt/pr-23813.
No original Verification log existed. Original linked evidence bundle and scripts were fetched with curl and read.
No real T3 data was read, copied, opened or seeded. No other machines used.

## Commands and observations

- git fetch origin; git worktree add /home/ubuntu/wt/pr-23813 <head>; vp install: success.
- agent-job was absent. Both compiler runs used (cd apps/server && vp run typecheck), invoking the repository-configured tsc. No tsgo or npx tsc.
- Downloaded probe-typecheck.sh adds six trivial methods to contracts, handlers, authorization and instrumentation. Local wrapper changes only its /tmp output directory and permits the deliberately replaced merge-base ws.ts. It restores all four files to HEAD on exit.
- Head: bash probe-local.sh <worktree> 6: exit 0, 204 Rpc.make declarations, 0 error diagnostics. head-summary.log and head-typecheck.log capture wrapper status and full compiler output.
- Merge-base comparison: git show <merge-base>:apps/server/src/ws.ts > apps/server/src/ws.ts; same +6 probe in the same installed worktree: exit 2, 20 TS377030 errors (requirements degraded to any). No TS2589 diagnostic was emitted in this rerun. base-summary.log and base-typecheck.log capture this. The historical 22-error claim refers to older evidence and is not the current count.
- Probe restored HEAD; git status --short was empty before runtime and after shutdown.
- Direct Node import of committed contracts: 198 RPCs, 104 split / 94 other, 198 unique tags, no overlap; every RPC carries middleware. partition-proof.json captures counts. Review A also checked every handler once against committed source.
- T3CODE_DEV_SEED=0 vp run dev:server --no-seed --home-dir /tmp/ev-23813/home --host 127.0.0.1 --port 15327.
- Startup confirmed baseDir=/tmp/ev-23813/home. Port-owner PID 806296 recorded once after listen, cwd /home/ubuntu/wt/pr-23813/apps/server; see server-owner.json.
- node apps/server/src/bin.ts auth pairing create --base-dir /tmp/ev-23813/home --scope orchestration:read --json: exit 0. Credential kept private; granted scope verified to be exactly orchestration:read.
- PROBE_ORIGIN=http://127.0.0.1:15327 PROBE_CREDENTIAL=<private file> PROBE_OUTPUT=runtime-proof.json vp exec node runtime-head.mjs: exit 0.
- Runtime probe is the downloaded probe plus a third call and assertions for guard rejection. Both partition calls succeeded on the same WebSocket: server.getSettings {} and scheduledTasks.list {}. server.updateSettings {patch:{}} failed with EnvironmentAuthorizationError, requiredPermission settings:write. HTTP token exchange and ticket both 200. Clean close 1000; no unexpected messages. See runtime-proof.json and runtime.log.
- Shutdown: checked recorded PID cwd again, SIGTERM to PID 806296; interrupted owned dev-runner session. Session exited 130; port 15327 no longer listening.
- Disk remained above 11 GB free. No commits, pushes, merges, comments, assignments, watches or browser use.

## Reviews and reconciliation

Fresh independent read-only reviewers ran in parallel on committed full diff vs merge-base. See reviewer-a.md (Codex gpt-6.1-sol) and reviewer-b.md (claudeAgent claude-opus-5-5). Both VERDICT: MERGE. No fallback needed.
Both traced full-group dispatch and authorization/instrumentation preservation, narrow Pick-based requirement extraction, and Effect toHandlers/toLayer implementation. Neither found a smaller materially better shape.
Reviewer B's approximate count is superseded by direct 104/94 counts. Its opening phrase "costs nothing at runtime" means no per-request work: there IS bounded linear construction/context-merge work per WebSocket layer. No benchmark claim is made. Finite headroom and possible future partition rebalancing remain disclosed.
The +60 and 318-method probes and six WebSocket tests remain historical evidence, not rerun in this round. They are not represented as fresh results.

Model/harness: proof audit and synthesis by gpt-6.1-sol through Codex in T3 Code; reviews by gpt-6.1-sol through Codex and claude-opus-5-5 through claudeAgent.

VERDICT: SUFFICIENT