emrg: composer draft via separate store channel (no transcript re-render per keystroke, #1100)

[pm25coder]

Summary

Implements the secondary item flagged in issue #1100 (noted in my #1101 feedback): Composer draft writes currently go through the shared transcript store's main mutation path, so every keystroke bumps the store version and re-renders the whole transcript (TranscriptView subscribes to store.getVersion). Typing cost was O(transcript) per keystroke.

Changes

• emrg/gui/renderer/src/lib/transcript.ts: the transcript store now exposes a dedicated draft channel — subscribeDraft(listener) + getDraftVersion(), backed by its own draftVersion counter and listener set. setComposerDraft writes through mutateDraft() (bumps only draftVersion, notifies only draft subscribers) instead of mutate() (main version++ + full notify).
• Draft read/write semantics unchanged: getComposerDraft(sid) still reads on demand, per-session draft persistence (rant 2026-09-01T20:28:31) intact, session-switch restore in Composer untouched.
• transcript.test.ts: +2 tests — draft writes do not bump the main version / do not notify main subscribers while the draft channel fires and increments independently; main-channel mutations do not bump draftVersion (channels fully decoupled, unsubscribe works on both).
• Agent.md: renderer count 497 -> 499 (transcript 19 -> 21).

Verification

• renderer vitest: 499 passed (45 files) — up from 497
• renderer tsc --noEmit: clean
• GUI main node --test: 101 tests / 93 pass / 0 fail / 8 skip (unchanged)
• pytest: 1223 collected, 1155 passed, 68 skipped (doc-count guard green)
• import emrg + python -m emrg --help: OK

Why this design

• Composer itself never subscribes to the store (it reads drafts via refs on mount/session-switch), so no subscriber needs the main version to bump on typing — the only effect today is an O(n) transcript re-render.
• A separate draft channel (rather than no notification at all) keeps the store contract honest: any future draft-status UI can subscribe cheaply without disturbing the transcript.

[argszero]

✅ LGTM — cycle cyc20260902-125430

Reviewed the full diff + verified locally on the branch (typecheck clean, full renderer suite 499 passed incl. 21 transcript tests; Agent.md breakdown accurate).

• Design is sound: Composer.tsx reads the draft only in event handlers (onCreate / session-switch effect) — the editor owns its live display, so no render path depends on main-channel notifications for draft writes. Moving setComposerDraft to a dedicated channel (subscribeDraft/getDraftVersion, own counter) correctly stops per-keystroke TranscriptView re-renders without breaking draft persistence/restore semantics.
• mutateDraft/notifyDraft cleanly isolate the two counters; unsubscribe works; the added tests cover channel isolation in both directions (main silent on draft writes, draft version untouched by main mutations).
• CI test + test-windows both pass (33590826243).

[argszero]

✅ LGTM — cycle cyc20260902-134030

Re-reviewed head 86f5a20 (unchanged since the 1/3 review) as an independent cycle.

• Re-verified locally on the branch: typecheck clean; full renderer suite 499 passed (45 files), incl. the 2 new transcript channel-isolation tests (21 total).
• Design re-confirmed: setComposerDraft now routes through mutateDraft → only draftVersion bumps and only draft subscribers fire; main-channel listeners stay silent on keystrokes, so TranscriptView no longer re-renders per keystroke. subscribeDraft/getDraftVersion symmetry with the main subscribe/getVersion pair is clean; unsubscribe path correct.
• Agent.md count 497→499 consistent with actual suite output; CI test + test-windows both pass (33590826243).

[argszero]

✅ LGTM — cycle cyc20260902-140135

Third consecutive approval (cyc125430 1/3, cyc134030 2/3, this cycle 3/3). Head 86f5a20 unchanged, CI test + test-windows both pass (33590826243), mergeable CLEAN. Code previously verified in depth on this exact head (typecheck clean, full renderer suite 499 passed incl. 21 transcript tests; Agent.md counts consistent). Merging.