[VarunGore36]

Five defects found while auditing the backend against its tests and contracts. Each commit is self-contained and comes with regression tests.

1. …Z timestamps were read as machine-local time (61dd17d)

generated_at / checked_at / enriched_at are written in UTC (time.strftime(…, time.gmtime())), but the three readers converted them with time.mktime(), which interprets the parsed struct as local time:

• app/services/poe_pipeline.py:is_stale — 180-day staleness badge
• app/services/marina_enrich.py:is_stale — 365-day enrichment window
• app/routers/formalities.py:_ts_of — auto-refresh cycle

On a machine at UTC+5:30 the windows slid 5 h 30 into the future (a zone generated today already looked stale); west of UTC they slide the other way (a two-day-old zone looks fresh and is never refreshed). calendar.timegm() reads the same strings as the UTC instants they are. routers/marinas.py already compared the UTC strings directly — the correct pattern was next door.

Tests: backend/tests/test_stale_timestamps_utc.py — exact epoch, staleness boundary ± 1 day, four timezones (UTC, Asia/Kolkata, America/Los_Angeles, Pacific/Kiritimati).

2. sitemap.xml inherited the host MIME table (36446d5)

The SPA route served build files through FileResponse without a media type, so the Content-Type came from the host's mimetypes database: application/xml on the CI runners, text/xml on machines whose /etc/mime.types lists xml that way (macOS). test_spa_static_head asserts application/xml for Search Console and failed there. The type is now pinned in app/main.py.

3. marina_build.haversine_nm raises ValueError: math domain error (5512dc7)

Third haversine in the repo, third time lucky: geo.haversine_km was fixed for near-antipodal points (a > 1.0 at floating-point rounding) and climatology_common.haversine_nm already guards asin(sqrt(a)), but this one did not. Same known trigger pair, same crash. One min(1.0, …), plus a regression test over both sibling implementations.

4. formalities.catalog_min_names was read but never catalogued (dbe99c8)

extract.py:catalog_is_sufficient resolves "enough names for a list without coordinates" via get_rule("formalities.catalog_min_names", 8) while the rule did not exist in backend/data/run_rules.json: not adjustable by a run, absent from every params.rules snapshot — against the catalogue principle ("a number has an interval and a why"). Entry added with the value already in use (8, interval 4–20).

5. test_memoire_grilles could not run on macOS (a180f89)

Its child process read /proc/self/status for an informational RSS printout and aborted before reaching it — the ceiling assertions (resident bytes, land mask, app import) are portable and ran either way. Now falls back to getrusage (ru_maxrss is bytes on macOS, KiB on Linux).

Verification

• backend: full pytest suite with the CI recipe (Mongo seeded, API up, requirements.txt env) — 1331 + 143 passed. The only remaining reds are 4 assertions on osm_confidence counts that require the CI's scripts/enrich_ci_osm.py Overpass step (data state, not code).
• frontend: 63/63 Jest, production build clean.
• infra/vps/ci (37), scripts/satellite (5), scripts/audit_tags: green.

Not in this PR

The macOS health probe (#40), the SearXNG default-port collision (#39) and the 40k LLM reader context (#37) are already proposed in their own PRs — deliberately not repeated here. The branch is based on main at d85f4d9 so the diff shows only these five fixes.