feat(update): add --watch mode for periodic re-indexing

[mavaali]

Summary

Adds qmd update --watch [--interval <dur>] [--embed] so users can keep the QMD index fresh without an external scheduler. This is the piece most often paired with qmd mcp in mcp.json-style setups where the agent expects the index to stay current as notes change.

Why

A friend of mine asked whether he could drop QMD into .vscode/mcp.json the same way you'd drop in @playwright/mcp and forget about it. Today the missing piece is re-indexing: qmd update works great manually, but there's no built-in way to keep it running. People end up wiring cron jobs, launchd agents, or just forgetting and getting stale results.

This PR adds watch mode so the common case ("re-index every few minutes, embed if anything new shows up") is one flag.

Design choices

• Polling, not fs.watch. reindexCollection() already hashes files and skips unchanged content, so idle ticks are near-zero cost. Polling avoids a new dependency (chokidar) and keeps cross-platform behavior identical on Linux, macOS, and Windows where fs.watch semantics differ.
• Default interval: 5 minutes. Configurable via --interval with units ms/s/m/h (e.g. 30s, 5m, 1.5h). Bare numbers are treated as seconds.
• Quiet by default. Ticks only emit a line when documents actually change. Routine no-op ticks are silent so the log stays useful. Format: [2026-05-15T20:27:31.844Z] smoke: 1 new, 0 updated, 0 removed.
• Optional --embed. Runs vectorIndex() after any tick that left pending hashes, so semantic search stays current without a second cron job.
• Circuit breaker. After 3 consecutive tick failures, the loop exits non-zero so a supervisor (launchd/systemd/Task Scheduler) can restart it instead of looping forever.
• Clean shutdown. SIGINT/SIGTERM stop the loop between ticks. Follows the same process.removeAllListeners pattern used by qmd mcp --http to bypass the top-level cursor-restoring handlers.

Usage

# Re-index every 5 minutes (default)
qmd update --watch

# Custom interval
qmd update --watch --interval 30s

# Auto-embed new hashes after each tick
qmd update --watch --embed

Paired with the MCP server:

# Terminal 1
qmd mcp --http --daemon
# Terminal 2
qmd update --watch --embed

What I didn't do

• Did not integrate into qmd mcp. Keeping watch as a sibling process avoids contention concerns and keeps the PR small. Happy to land that as a follow-up if you'd like.
• Did not add a launchd/systemd template. Documented the pattern in README; users can wire their own supervisor.
• Did not use fs.watch. Open to revisiting if anyone wants true real-time, but polling covers the stated use case.

Tests

• 17 new unit tests in test/watch.test.ts covering parseDurationMs (units, fractions, edge cases, error paths) and runWatchLoop (tick counting, failure tolerance, circuit breaker, consecutive-failure reset, invalid interval).
• All 843 existing tests still pass.
• E2E manual smoke test: created a fresh collection, started qmd update --watch --interval 1s, added a new .md file mid-run, observed exactly one timestamped log line, sent SIGINT, observed clean exit.

Refactor notes

updateCollections() now takes { quiet?, keepDbOpen? } and returns an UpdateSummary. Non-watch behavior is byte-identical to before. The only externally visible change in the legacy path: the existing console.log for custom-update-command stderr now routes through console.error (correct destination for stderr content; previously it went to stdout).

Checklist

• Tests added
• npm run build passes
• Full test suite passes (npx vitest run)
• Changelog entry under ## [Unreleased]
• README updated