fix(web): reflect on idle so web conversations evolve Lisa's desires

[oratis]

Stacked series (1/3). This is the base of a 3-PR stack fixing why desires don't evolve. Order: #242 → #246 → #247. #246 and #247 are branched on top of this one.

Why

Investigating "the wish never changes no matter what we talk about" traced to a dead code path: reflectOnSession — the only automatic way a conversation turns into or updates a desire — is never triggered in the web deployment. It's wired to CLI --reflect and the channel router only; the POST /reflect endpoint exists but no client calls it. So web conversations never touched Lisa's desires. This is the dominant root cause (full analysis in the plan doc).

What

A dedicated, debounced reflection scheduler in startWebServer:

• After LISA_REFLECT_DEBOUNCE_MS (default 5 min) of no user input, reflect once on the accumulated history.
• Only when the user added a new message since the last reflection — counting user messages (not total) so Lisa's own idle [while you were away] notes never trigger reflection on her own monologue.
• Independent of the 60-min "dream" idle (src/cli.ts default) — works even with --no-idle. It reuses the process-wide idle watcher purely as an activity clock (/chat already tick()s it).
• Robust: advances its marker only on success (a failed reflect retries next tick); snapshots history before the async call; unref()s its timer; skips while a reflection or idle run is in flight.
• Emits an SSE reflect_done event and appends to the session log for observability.

Decision logic is extracted to a pure, unit-tested module (src/web/reflect-scheduler.ts).

Plan doc

This is PR 1 of 3 from docs/PLAN_DESIRE_EVOLUTION_v1.0.md, which bundles the root-cause analysis, the pro/con debates (正反方辩论), and the PR1–3 breakdown. PR2 makes reflection able to revise/close desires (not just append); PR3 surfaces the most recently active desire instead of a filesystem-order pick.

Testing

• src/web/reflect-scheduler.test.ts — 8 tests covering the debounce/guard truth table and user-message counting. All green.
• npm run typecheck: no new errors (pre-existing imapflow optional-dep error only).
• npm test: zero regressions — same 5 pre-existing failures as main (imapflow mail tests, a stale HTML snapshot, an auth test), plus the 8 new passing tests.

Backend timer change — not observable in the browser preview; verified via unit tests + typecheck.

🤖 Generated with Claude Code

[oratis]

Review — merge-ready ✅

Walked the concurrency + debounce truth table against the code:

• --no-idle really works (the central claim) — /chat calls getIdleWatcher(...).tick() unconditionally, so the scheduler's activity clock advances even when the 60-min idle block is skipped. The scheduler only uses idleFor() + its own interval; it never needs the watcher's idle event.
• History snapshot + snapshotUserCount + reflecting = true are set in one synchronous block (no await between), and the marker advances to the snapshot count only on success — so a message arriving mid-reflect is counted exactly once next tick, never lost or double-counted.
• reflecting is set synchronously before the async IIFE and reset in finally (even on a sync throw), so two reflects can't overlap; a failed reflect retries without advancing the marker.
• Soul writes serialize on the cross-process file locks and session-log appends are O_APPEND atomic, so a reflect + the dream can't corrupt shared state.

Sound — no blockers. Two small optional hardenings I'll fold in while merging:

1. Clamp LISA_REFLECT_DEBOUNCE_MS to > 0. Number(env) || DEFAULT correctly rejects NaN/""/"0", but a negative value is truthy → debounceMs = -1 makes idleMs < debounceMs always false, so it reflects ~60s into any conversation. Number.isFinite(n) && n > 0 ? n : DEFAULT.
2. Have the 60-min dream also respect reflecting. The scheduler's inFlight: reflecting || idleRunning blocks reflect-during-dream, but the dream's guard is only if (idleRunning) return — so under a short --idle (both firing on the same quiet trigger) a dream can start mid-reflect and its own [while you were away] note can bleed into the reflected history — the exact thing PLAN §3 wants to avoid. Adding || reflecting to the dream guard closes the asymmetry.

Neither blocks the fix; merging after those.