feat(vitals): per-version crash/ANR breakdown + stalled-feed warning

[TimeToBuildBob]

Two gaps surfaced by trying to answer "are crash rates better on v0.14.0b2?" (ActivityWatch/aw-android#176). Neither is answerable with the tooling as it stands.

1. The app-wide series can't attribute a rate to a release

data/android-crash-rate.csv is one userPerceivedCrashRate number per day across all installs. During a partial rollout that number is dominated by the install base still on the old build, so a genuinely fixed release barely moves the aggregate for weeks — and when it does move you can't tell the release from seasonality.

Adds vitals.py by-version, which requests the same metric with the API's versionCode dimension and compares versions directly:

uv run vitals.py by-version --days 14
uv run vitals.py by-version --kind anr-rate --as-csv

2. A stalled feed was invisible

_daily_freshness anchors each query to the API's latest available DAILY date. When that date stops advancing, the collector still exits 0 and rewrites identical rows — so git diff --cached --quiet skips the commit and the daily job keeps reporting success while the series silently stops.

That is not hypothetical: android-crash-rate.csv and android-anr-rate.csv were last updated 2026-08-22 (data through 08-19) while installed.csv updated today and every collect-play.yml run in that window is green.

_freshness_warning now prints the lag to stderr past STALE_FRESHNESS_DAYS (3), so a stall shows up as output rather than as absence of it. It does not fail the job — diagnosing whether this stall is Play-side or permissions-side needs the service account, which I don't have.

Notes

• by-version prefers the API's valueLabel when it adds information, so rows read 1000 (0.14.0b2) rather than a bare code.
• No behaviour change to the existing crash-rate / anr-rate / summary commands or the collected CSV format.
• I have no Play service-account key, so the live API path is unverified; the query shape is covered by --dry-run and the parsing/warning logic by offline tests.

Verification

uv run ruff check ., uv run mypy *.py, and uv run pytest -q (11 passed) all green locally — 9 new tests in test_vitals.py covering the dimensioned query body, per-version row parsing, and the freshness warning.

[greptile-apps]

Greptile Summary

The PR adds per-version crash and ANR reporting and warns when Play’s daily vitals feed becomes stale.

• Adds dimension-aware vitals queries and a by-version CLI command.
• Groups results by raw version code while retaining labels for display.
• Uses csv.writer for correctly escaped CSV output.
• Adds offline coverage for query construction, parsing, grouping, CSV encoding, and freshness warnings.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the current implementation uses the raw version code as its grouping identity and delegates CSV field encoding to csv.writer.

Important Files Changed

Filename	Overview
vitals.py	Adds dimensioned reporting and stale-feed warnings; both previously reported issues are addressed at the current head.
test_vitals.py	Adds focused offline tests for the new query, parsing, grouping, CSV, and freshness behavior.

_{Reviews (2): Last reviewed commit: "fix(tests): split imports and remove unu..." | Re-trigger Greptile}

[TimeToBuildBob]

Fixed both Greptile findings in e50c0da (16 tests pass):

• P1 stable version identity: added _dimension_raw as the grouping key (raw versionCode only, no label), version_labels dict tracks the best display string seen. Same versionCode with intermittent missing labels now lands in one bucket.
• P2 CSV escaping: --as-csv now writes through csv.writer; commas/quotes/newlines in Play version labels are properly encoded.

[TimeToBuildBob]

CI lint failure fixed in 617a113: split import csv, io as _io onto separate lines and removed the now-unused import io above it. 16 tests still passing.

[TimeToBuildBob]

CI is green, both Greptile findings (P1 version-identity grouping, P2 CSV escaping) are fixed. Ready for maintainer review and merge.

[TimeToBuildBob]

@greptileai review

[TimeToBuildBob]

Current-head re-review is complete at 5/5 with no blocking findings; CI is green. I do not have merge permission on ActivityWatch/stats, so the only remaining action is maintainer review/merge.