Delete `terminals:dirty` channel; collapse session autosave to direct write

[srid]

Symptom

packages/server/src/session.ts:89-108 runs an initSessionAutoSave loop that subscribes to the terminals:dirty channel, leading-edge-throttles at 500ms, and persists the session blob.

The channel fires at high frequency: packages/server/src/meta/state.ts:63 publishes terminals:dirty on every metadata publish, including live-only fields. Per the comment at session.ts:80-84, the Claude transcript watcher fires terminals:dirty every 150ms while an agent is streaming.

But SavedTerminal (packages/common/src/index.ts:152-176) persists only cwd, git, themeName, parentId, canvasLayout, subPanel, lastAgentCommand. The high-frequency live fields (agent, pr, foreground, declared in LiveTerminalFieldsSchema at index.ts:184-191) are not persisted.

Result: during agent streaming, terminals:dirty fires ~6×/sec, the throttle schedules a save every 500ms, and that save writes a SavedTerminal blob whose persisted bytes are identical to the previous one. Most saves during streaming are no-op identical writes that still publish to every connected client via session:changed.

Why this matters

• session.ts carries the only autosave-with-throttle in the codebase. preferences.ts (packages/server/src/preferences.ts:28-40) writes directly at the mutation site. Session is the odd one — because the signal is overgenerous, not because the content churns.
• The throttle subscription owns a module-level saveTimer and a documented test-ordering race (session.ts:55-65) that exists only because the throttle exists. Both vanish under the fix.
• Disk and publishSystem("session:changed", …) traffic during agent streaming is wasted.

The fix is simpler than first sketched: delete the channel

Initial sketch was "split terminals:dirty into live vs. persisted-dirty channels." Investigation showed this is overcomplicated. Grep for terminals:dirty returns:

• 2 publishers: terminals.ts:78 (create/kill), meta/state.ts:63 (every metadata publish).
• 1 subscriber: session.ts:97 (the autosave loop).
• Zero other consumers.

The live UI does not use terminals:dirty. Live terminal.list rides on the terminal-list channel (publisher.ts:46-47); per-terminal metadata streams ride on metadata:<id> channels (publisher.ts:21-23, published at meta/state.ts:62). terminals:dirty is a 1:1 pipe whose entire purpose is to be the throttle's input. There's nothing to split.

Fix shape

session.ts collapses to the preferences pattern: saveSession(snapshot()) is called directly at the small set of sites that mutate persisted fields:

• onCwd callback in terminals.ts:135 (writes cwd via updateServerMetadata).
• The git provider when git resolves/changes (packages/server/src/meta/git.ts).
• Client-driven RPCs that write themeName / canvasLayout / subPanel / parentId.
• The preexec hook that writes lastAgentCommand (packages/server/src/meta/agent-command.ts).
• createTerminal (terminals.ts:92-172) — initial seeding.
• killTerminal and the onExit callback (terminals.ts:108-122, terminals.ts:175+) — removal.

Once those direct calls land, terminals:dirty has zero consumers. Delete:

• The channel typing in publisher.ts:45 ("terminals:dirty": Record<string, never>).
• Both publishers: terminals.ts:78, meta/state.ts:63.
• The subscription loop and saveTimer in session.ts:89-108.
• The cancellation logic and the test-ordering race comment in session.ts:55-65 (setSavedSession's clearTimeout block).

Result: three persistence modules (preferences.ts, session.ts, future tasks.ts) share one shape — direct store.set at the mutation site + a single :changed publish — and the codebase loses one channel, one timer, and one documented race.

Risk

store.set is sync and cheap, but publishSystem("session:changed", …) reaches every connected client. Worst case under the fix is N writes for N rapid persisted-field mutations (e.g. session-restore replay) — vs. one write per 500ms today. Audit the batched-mutation sites before merging; if a same-tick burst is meaningful, a small runWithSave(() => { … saveSession() }) wrapper at each call site is sufficient (no need for a publisher-throttled loop).

Out of scope

tasks.ts (the new persisted domain landing under #760's refinement) does not depend on this fix; it joins the cluster directly with the preferences pattern. This issue is about cleaning up the fourth module to match.