[Juliusolsson05]

Fixes #1117

What was wrong

When opencode db path fails at launch, the package retries it (#1114). When it later succeeds, DurableReader.position() seeds from the database's current head, so every row the TUI committed while the pane was dark counts as already seen. Those rows are never emitted, and the pane shows an answer whose prompt never appeared, or nothing at all. The package reports db_path_recovered_late ("Reload the session to re-read it") because it cannot tell a dark-window row from last week's.

Fix

useIpcSubscriptions.ts answers db_path_recovered_late by running loadInitialHistoryForSession once. A history read admits only uuids the pane does not already hold, and placeHistoryEntries places them in order around the rows it does hold. On success the status becomes ready; if the load fails, the error stays, which is still true. The diagnostic never reaches the lifetime banner (unchanged).

Evidence

The OpenCode Terminal harness gains a lateDatabase option. It launches dark while the replay keeps committing to the real fixture database, and uses a mocked resolveOpencodeDbPath that starts answering on demand. The new test in opencodeTerminalDiagnostics.renderer.test.tsx runs through the real package, SessionManager and forwarder with the recorded plain.json:

• Fail-first: without the fix, it failed with timed out waiting for the dark-window prompt, after the recovery had fired.
• With the fix: the recorded prompt is in runtime.entries, the status is ready, and there is no lifetime banner.
• Existing banner test: it asserted transcriptError for both codes. For recovered_late that value now belongs to the heal, so it asserts only the banner, with a comment explaining why.
• Suites: renderer 3,034 of 3,034 pass (Node 24). tsc -b is clean.

Not verified

Live OpenCode (house rule). The package's own dark-window test (#6) is unchanged.

🤖 Generated with Claude Code

Review disposition (A Grok, B Pi; 2 rounds)

Finding	Disposition
A MAJOR: the late open's first store open BUSY → retry via openDurable never emitted db_path_recovered_late, so no heal ran and the pane looked healthy	Fixed in the package (opencode-terminal-headless#9, merged; submodule at its merge commit 6f343e7c): the report fires wherever the reader starts. New package system test
A/B: a failed heal's error was cleared by the next committed row	Fixed (1f23a5d8): a failed heal sets the recovered-late message as the lifetime banner
B round 2: that banner then survived a later SUCCESSFUL read, so parents saw transcript_unavailable over a complete transcript	Fixed (d52a779b): a successful history read clears that one banner (channelErrorAfterRead); other channel errors stand. The heal failure no longer resurrects a closed pane's runtime. Test extended; the mutant is killed
B1: the recovered-late write was unpinned	Pinned: the heal ran exactly once; session:jsonl-error reached the pane
B3/B4, A suspicion: the history source and the pane resolve the DB path with separate memo keys	Noted: for an absolute binary the keys coincide on the normal boot path; no change

Round 2: A and B both MERGE-READY. The round-2 banner fix followed, with a test and a mutation check.