feat(calendar): add provider-neutral availability tool

[boseq]

Summary

• add a read-only get_calendar_availability tool over the shared local calendar cache
• combine selected Google, Microsoft, and Apple calendars into privacy-safe busy intervals and available slots
• normalize provider free/busy and self-declined semantics, including all-day events, overlap boundaries, buffers, and minimum slot duration
• add IPC/preload wiring, prompt guidance, UI status/icon support, types, translations, and focused coverage

Scope

• no new OAuth scopes, token flows, or provider connections
• no raw event titles, attendees, meeting links, account identifiers, or provider names are returned to the model
• results are explicitly labeled as local-calendar-cache coverage for the next seven local calendar days

Validation

• 45 focused calendar availability/provider/database/tool tests
• TypeScript typecheck
• full ESLint
• i18n consistency check
• Swift typecheck
• renderer production build
• Prettier and git diff checks

Manual testing

Validated locally with Google, Microsoft, and Apple events in a 9 AM–6 PM range, including minimum-duration filtering and cross-provider busy interval union.

[Chadpiha]

Approving — solid work. I did a full pass on this: read every changed file, traced the flow end-to-end (prompt context → tool call → renderer validation → IPC → service → cache query → interval math → localized facts back to the model), and ran the suite, typecheck, i18n check, and a renderer production build locally.

Things I specifically verified:

• Privacy boundary holds. Only times, durations, and flags ever reach the model; the projection fails closed on malformed payloads, and unknown IPC errors collapse to a generic message with an allowlist for the correctable ones.
• Cache coverage vs. the 7-day horizon checks out for all three providers (Google 9d lookahead + 1d token TTL, Microsoft 14d/7d TTL + lookback, Apple 8d snapshot + 2h lookback), including ongoing multi-day events since all three provider queries are overlap/end-time based.
• Microsoft all-day date-only storage — I confirmed against Graph's documented behavior that all-day events stay date-pure at midnight in the preferred zone, and the format change can't regress reminders or the upcoming UI because those queries exclude is_all_day rows.
• Migration safety — table ordering is right, fresh DBs no-op, and the forced REST resync populates the new columns; until then unknown conservatively blocks, which is the right failure direction.

I pushed two small commits onto the branch:

1. fix(calendar): load shared availability helper as ESM — the constants-sharing commit imported a CJS (module.exports) helper into the renderer graph, which hangs the Vite-SSR chat-streaming tests (that's why the CI tests jobs were stuck). Converted the helper to ESM per the meetingJoinUrl.js convention, and split computeCalendarAvailability out so the service validates each request once.
2. fix(calendar): accept null optional args in availability tool — models often send explicit null for optional params; the renderer now treats null as absent like the service's ?? defaults already did (with a regression test), and the seven-day wording in the schema is now derived from MAX_AVAILABILITY_HORIZON_DAYS.

Two behavioral notes for the record, both fine by me: unanswered invitations block availability on all providers (only an explicit decline or free/transparent unblocks — matches provider free/busy semantics), and broad multi-day questions intentionally get a clarifying question about daily hours before per-day calls.