fix(claude): explain missing-transcript resume failures; add t3 session audit|reset

[DanielGGordon]

Problem

Sending in a Claude thread whose persisted resume cursor points at a transcript that no longer exists on this host failed with only "Claude runtime stream failed." The real cause — Claude Code's No conversation found with session ID: <id> — was emitted one event earlier and then overwritten by the generic stream-exit error, so the thread gave no hint at what was missing or how to recover. Found on prod thread 89de3af4… ("Second Brain Architecture Review"); an audit of the prod DB shows 72 non-deleted Claude threads (43 unarchived, mostly claude-import-* from another machine) in the same state.

Changes

Adapter (ClaudeAdapter.ts, claudeSessionTranscript.ts)

• Recognise the CLI's missing-transcript result error and emit a runtime error that names the session id, the expected ~/.claude/projects/<encoded cwd>/<id>.jsonl path, and the two ways out (restore the file and resend, or t3 session reset <threadId>). Detail payload carries reason: "resume_transcript_missing".
• Keep that explanation through the subsequent stream exit instead of replacing it with "Claude runtime stream failed."
• The resume cursor is deliberately not cleared — the transcript may still be recoverable from the machine the conversation ran on.

CLI (cli/session.ts, shared cli/offlineRuntime.ts)

• t3 session audit [--json] [--include-deleted] [--all] [--projects-dir] — lists every Claude thread whose transcript is missing (or relocated under a different project folder) with lifecycle, title, cwd, session id and the exact path to restore. Read-only.
• t3 session reset <threadId> [--yes] — prints the current cursor (so it can be restored by hand) and only clears it with --yes.
• Both run offline against the state DB and are safe while the server is serving (WAL; the server re-reads bindings on every session start). The offline runtime layer that was private to t3 import is now shared.

Docs: README feature bullet, AGENTS.md ops section.

Verification

• tsgo --noEmit clean; vp lint clean; vp fmt --check clean.
• New tests: claudeSessionTranscript.test.ts (encoding, error parsing, message), cli/session.test.ts (binding → target, lifecycle, row formatting), ClaudeAdapter.test.ts ("explains a resume whose transcript is missing and keeps that explanation on stream exit").
• Full apps/server suite: 1777 passed; the single failure (server.test.ts > preserves structured workspace rpc failures) is pre-existing and fails identically on main.
• Ran t3 session audit against the live prod DB: 137 Claude threads with cursors → 65 present, 0 relocated, 72 missing, ~4s. t3 session reset dry run verified to leave the row untouched; on a thread with no cursor it reports "nothing to reset".

🤖 Generated with Claude Code

[DanielGGordon]

Test deployment

This branch is running on test slot 7448 with a copy of the prod threads (so the stale-cursor threads are there to try).

Pairing link: https://15.204.108.12:7448/pair#token=9DHXVHVUCX8R
(one-time, short expiry — if it's stale, re-mint on the server with node scripts/test-status.ts --pair 7448)

To see the change: open a Claude thread whose transcript is missing (e.g. "Install Cursor CLI" or any claude-import-… thread for DanCode / gphotos-to-constant-contact) and send anything. Instead of "Claude runtime stream failed." the thread now shows which ~/.claude/projects/…/<session>.jsonl is missing and the two ways out (restore the file, or t3 session reset <threadId>).

CLI, from the worktree on the server:

node apps/server/src/bin.ts session audit            # read-only list of affected threads + paths to restore
node apps/server/src/bin.ts session reset <threadId> # dry run; add --yes to actually clear the cursor

Slot: external 7448 ⇄ loopback 3778.