fix(ingest): convert tachometer parquet that lacks timestamp_ns

[nv-yna]

What

src/ingest/metrics_tachometer.py now converts tachometer parquet that has no timestamp_ns column. When the column is absent, the epoch of time_since_start == 0 is taken from, in order:

1. an explicit --start-ns (new CLI flag; process(..., start_ns=)),
2. the first timestamped line of the run's tachometer.out (searched up to four levels above the parquet; the scraper logs within the second it starts counting from),
3. the newest parquet's mtime minus its largest time_since_start (row-group statistics, no data read).

The source used is logged at WARNING. Precision equals the one-second grid every timestamp is snapped to anyway. Parquet that carries the real column is processed exactly as before.

Why

Every published tachometer-scraper release asset up to v1.0.96 predates the vendored writer's timestamp_ns column (#350, merged 2026-08-26). The release workflow's build-tachometer-scraper jobs have failed at actions/checkout@v4 on every merge that touched the Rust sources since then (runs 33043846919 for #350 and 34112274235 for #401, both jobs failing within 4 s), the release job was skipped, and every later release took the "reuse scraper binaries from the previous release" path. Result: the v1.0.96 aarch64 asset is byte-identical (sha256 344181ae…) to a download from 2026-08-27, and make setup installs it.

Such scrapers write only time_since_start, and the ingest died on every capture:

File ".../src/ingest/metrics_tachometer.py", line 226, in process
    tss = batch.column("timestamp_ns").to_pylist()
KeyError: 'Field "timestamp_ns" does not exist in schema'

Seen on hecate job 487539 (2026-08-27) and again on 565810 (2026-09-09). The metrics leg of the perf dashboard bundle was therefore silently empty for runs that had the full capture on disk.

The release workflow itself needs a separate fix (the pull_request_target checkout of merge_commit_sha); until the assets are refreshed, make tachometer-scraper (or the Dockerfile cross-build) produces a correct binary. This change makes the ingest independent of which binary a run happened to use.

Why draft

Held until the three hecate runs that exercise it finish and post-process end to end: 565810 (old binary, fallback path) and 565811 / 565854 (rebuilt binary, timestamp_ns present). Will attach the ingest log lines from each before marking ready for review.

Validation

• tests/test_metrics_tachometer.py: 4 new tests (log-anchored, mtime-anchored, explicit --start-ns, and real-column-preferred-over-stale-log) + 8 existing → 12 passed; pytest -k "tachometer or ingest or metrics" → 102 passed.
• Real data: job 565810's live legacy parquet (11,007,786 rows, 17 min of capture) → 1,010 one-second timestamps 2026-09-10T00:37:10Z..00:54:02Z, anchored on the scraper's first log stamp 00:37:08Z; 125 families in the first line. ~2.5 min on the login node.
• Lint: ruff check on the two files reports the same 5 findings main already has (RUF100, SIM108, SIM115, RUF046 on the pre-existing snap line); this diff adds none.

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@b943e59). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #414   +/-   ##
=======================================
  Coverage        ?   75.05%           
=======================================
  Files           ?      109           
  Lines           ?    16002           
  Branches        ?        0           
=======================================
  Hits            ?    12011           
  Misses          ?     3991           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[nv-yna]

Evidence from the three hecate runs (2026-09-09, AgentX baseline recipe, 8 VR200 nodes)

All three postprocessed with this branch's metrics_tachometer.py (cherry-picked onto the run checkouts). Lines are from outputs/<job>/logs/sweep_<job>.log.

565810 — release-asset scraper (v1.0.96 binary, no timestamp_ns column) → fallback path taken:

19:21:28 INFO [L2 metrics] auto-selected source: tachometer
19:21:28 INFO [L1] metrics raw: .../565810/logs/tachometer/raw/scrape/incomplete-49.parquet (tachometer parquet)
19:21:29 WARNING timestamp_ns absent from the tachometer parquet (scraper predates #350); deriving timestamps from time_since_start + 2026-09-10T00:37:08+00:00 (.../565810/logs/tachometer.out)
19:30:31 INFO [L2 metrics] tachometer -> server_metrics_export.jsonl: 6106 timestamps, dedup 6106 -> 6106 lines
19:32:15 INFO [done] bundle ready in 649.3s: aiperf=True traces=False metrics=True request_trace=False

Without this change the same step raised KeyError: 'timestamp_ns' and the bundle had no metrics (the failure that started this PR). Parquet inventory for that run: incomplete-49.parquet | rows: 70048941 | has timestamp_ns: False.

565811 and 565854 — scraper rebuilt from source (has timestamp_ns) → unchanged fast path, no warning:

21:19:44 INFO [L1] metrics raw: .../565811/logs/tachometer/raw/scrape/incomplete-21.parquet (tachometer parquet)
21:22:53 INFO [L2 metrics] tachometer -> server_metrics_export.jsonl: 5661 timestamps, dedup 5661 -> 5661 lines

22:02:19 INFO [L1] metrics raw: .../565854/logs/tachometer/raw/scrape/incomplete-50.parquet (tachometer parquet)
22:10:05 INFO [L2 metrics] tachometer -> server_metrics_export.jsonl: 6319 timestamps, dedup 6319 -> 6319 lines
22:11:45 INFO [done] bundle ready in 568.5s: aiperf=True traces=False metrics=True request_trace=False

(incomplete-21.parquet | rows: 21008645 | has timestamp_ns: True.)

Two adjacent observations, not fixed here: the scraper never writes final.parquet (no shutdown compaction), so ingest always reads the last incomplete-N shard; and every release asset up to v1.0.96 is the pre-#350 binary because release.yaml reuses the previous release's tachometer binaries when the merged PR does not touch Rust sources and the #350 rebuild failed at actions/checkout. This fallback keeps those deployments producing a metrics bundle until the release pipeline ships a current binary.