feat(stats): add --by-repo and --by-time breakdowns to teamai stats

[Eyre921]

Summary

teamai stats reported tokens / turns / interventions as a single lump. This adds two optional breakdowns over the same local dashboard event stream — no new data collection:

teamai stats --by-repo   # per-project: sessions / turns / tools / tokens / interventions
teamai stats --by-time   # 24h activity sparkline, peak hour, active minutes, night-owl %

Ports the shape of claude-cloud-sync's stats.json analytics (by_repo, timeline_by_hour, active_minutes, night-owl ratio) to the data teamai-cli already has.

⚠️ Stacked on #191 and #192. This PR's own change is the final commit; please merge those first — the diff will then reduce to just the stats change.

Type of Change

• New feature (non-breaking change that adds functionality)

Details

• Attribution granularity: per session (one cwd per session), not the sync's per-turn cost splitting — that's what the event stream carries. Honest about the trade-off.
• utils/repo-attribution.ts — attributeRepo(cwd) maps a working directory to a project label (basename), and canonicalRepo() ports repo_canon so a remote-form github.com/o/r and cnb.cool/o/r collapse to the same platform-independent owner/repo. Today a filesystem cwd degrades to its basename; the canonicalization is future-proof for when a git remote is captured.
• session-analytics.ts — attributeByRepo() and timeAnalytics() are pure folds; timeAnalytics uses the machine's local timezone for hour-of-day.
• Default teamai stats output is unchanged; the new sections appear only with the flags.

Test Plan

• npx tsc --noEmit
• npx vitest run — 141 files, 1754 tests (13 new in session-analytics.test.ts)
• npm run build
• Real-CLI E2E: seeded ~/.teamai/dashboard/events.jsonl across two repos and two hours; node dist/index.js stats --by-repo --by-time printed a per-repo table (alpha 150 tok / beta 15 tok, sorted) and an hourly sparkline (peak 02:00, active 4 min, night-owl 67%) — all matching the seed.

Notes for Reviewers

• no_repo is used for home/root/ops cwds; the per-repo table includes it when present.
• Timezone: hours are local to the machine running stats (the sync fixed this to Beijing time; local felt more appropriate for a per-user CLI). Easy to make configurable if preferred.

[jeff-r2026]

Suggestion: make this PR independent of #191/#192

I confirmed that the stats change itself is functionally standalone — the stacking here is purely at the git level, not in the code.

The actual commit for this PR (6cc98eff) touches only 5 files, and none of them import anything from #191/#192:

File	Imports
src/utils/repo-attribution.ts	none (zero-dependency util)
src/session-analytics.ts	only dashboard-collector, types, repo-attribution
src/stats.ts	only adds ./session-analytics — no redact, no session-collector
src/index.ts	only the --by-repo / --by-time options

The redact / save-session / session-collector code shows up in the diff solely because the three commits sit on the same branch (redact → save-session → stats, with stats on top), so gh pr diff renders all three.

Could you rebase this so it stands alone? Cherry-picking 6cc98eff onto a fresh branch from origin/main will collapse the diff to just the 5-file stats change, and it can then be reviewed and merged without waiting on #191/#192.

git fetch origin
git switch -c feat/stats-by-repo-only origin/main
git cherry-pick 6cc98eff
# force-push to the PR branch, or open a new PR against origin/main

Independent of the split, three small follow-ups on the stats code:

1. renderHourSparkline renders 0 and the lowest non-zero hour as the same ▁ bar — consider a blank/· for empty hours so idle vs. low activity are distinguishable.
2. --by-repo silently truncates to top 15 — add a … and N more line when there are more.
3. Add a direct unit test for renderHourSparkline (all-zero + single-peak cases).

[Eyre921]

Rebased to stand alone — cherry-picked onto a fresh branch from origin/main, so the diff is now just the 5-file stats change in a single commit (no #191/#192 code). Thanks for confirming it was functionally standalone.

All three follow-ups done:

1. renderHourSparkline now renders empty hours as ·, visually distinct from ▁ low-but-nonzero activity.
2. --by-repo prints … and N more when there are over 15 repos.
3. Added direct unit tests for renderHourSparkline (all-zero → 24 dots; single-peak → one █ + dots).

Re-verified: tsc --noEmit, vitest run (142 files / 1760 tests, 15 in session-analytics.test.ts), npm run build, and a real-CLI E2E showing the … and 2 more truncation and the dotted idle hours.

Note: the CI DSN-scanner finding is not from this PR's files, but the whole-repo scan may still flag it until #207 (the redact.test.ts fixture fix) merges.