355615d27edd0f55055e07666caa6dea04f3f642
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 builtbykev
|
e5b20b0509 |
Software watches coverage now, and ancestry becomes a product contract
Three closeout items, traced before building.
THE OBSERVER WAS DIAGNOSTICS, NOT MONITORING. Traced from the deployed tree:
exactly two callsites, both manual internal routes, and nothing in the scheduler
or ops path consumed it. A persistent lineage failure could have sat unnoticed
until a human asked.
The monitor now runs on the scheduler's per-minute tick, throttled to 30
minutes. It is placed there rather than after a snapshot on purpose: an
in-snapshot audit structurally cannot report that no snapshot ran, which is the
failure mode that matters most, and it would run under peak write contention
where the audit already demonstrably times out. It reads only the durable
observer and never `attachLineage`'s counters, and every failure path is
swallowed — a monitor that can take down the pipeline it watches is worse than
no monitor.
HEALTHY IS SILENCE; EVERYTHING ELSE SPEAKS. `coverageAlarm` is pure, so the
policy is testable and cannot drift into the scheduler. AUDIT_UNAVAILABLE says
"could not be measured — the audit did not run", deliberately worded so it can
never be read as "coverage is zero": those are different claims and collapsing
them is how a monitor starts lying in the reassuring direction. Alerts dedupe on
(health, cohort) so a standing fault states itself once and a NEW cohort with
the same fault speaks again.
ANCESTRY BECOMES A PRODUCT CONTRACT. It was internal-only. `GET
/api/ancestry/ledger/:id` (requireAuth, rate-limited) plus the Next proxy that
makes it browser-reachable, keyed on the LEDGER ROW ID — a stable identifier the
ledger API already returns — rather than a raw natural key exposed because it
was convenient. Its own router, so `routes/ledger.js` stays free of lineage
entirely and the grade-badge guard keeps its teeth. Every response declares
`authority: LOCAL, authority_scope: ANCESTRY_ONLY`.
THREE TEETH CAME BACK GREEN AND ALL THREE WERE MY TESTS, NOT SAFE DEFECTS.
The badge guard was CASE-SENSITIVE, so `LINEAGE_ANCESTRY` and `readAncestry`
walked straight past it. `try/finally` is valid JavaScript, so removing the
monitor's catch produced no load error and nothing asserted the containment.
And the multi-date cohort check was a grep for `.gte('game_date'` that the
head query satisfied on its own. All three replaced with behavioural tests,
including a scheduler double whose fake client HONOURS its filters — a
pass-through would have made a narrowed cohort walk look correct.
Suite 398/5,522/0 · tsc 0 · web build 0 · teeth 14/14 and 15/15.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CQJeAG8vcDoL5zkiaJyVb8
|
||
|
builtbykev
|
1728ce4fac |
A cohort is a snapshot_id, not a date slice of one
Found by cross-checking the observer against an independent SQL derivation of the same natural cohort instead of accepting its own answer. The 2026-08-31 03:00Z run held 1,970 rows across TWO game_dates — 1,572 on 08-31 and 398 on 08-30, the late slot spanning midnight ET. The observer selected the head row's date and filtered to it, so it measured 646 of the 801 eligible keys and reported HEALTHY. Both slices happened to be fully covered, so the verdict was right by luck; a date slice with zero lineage would have been invisible behind it. The audit now covers the whole snapshot_id and reports every game_date it spans. A test builds the real two-date shape with one slice uncovered and requires PARTIAL_COVERAGE. Suite 397/5,499/0 · 15/15 teeth. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQJeAG8vcDoL5zkiaJyVb8 |
||
|
builtbykev
|
1bdc3baec7 |
A valid chain can still have a hole in it
Measured against real production history while validating the ancestry surface: `aj ewing / total_bases / over / 0.5` on 2026-08-28 carries a valid ORIGIN from the 18:10 cohort AND a row from the 19:03 cohort that was stamped for publication and never recorded an action. The chain is genuinely valid and the partial row is correctly excluded from it — that part of the contract held. But answering LINEAGE_AVAILABLE and nothing else presents an incomplete chronology as a complete one, which is the exact failure this surface exists to make visible, one grain finer than a whole Read. The response now carries `chronology_complete` and `unrecorded_publications`. The state is unchanged, because the chain that exists really is valid; what changes is that the gap is disclosed instead of smoothed over. Suite 397/5,498/0. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQJeAG8vcDoL5zkiaJyVb8 |
||
|
builtbykev
|
930d526b01 |
Lineage productization: a live graph, an observer that doesn't trust it, and a surface that owns only what it knows
The authority review found there was nothing to switch: lineage is a
write-only graph with no living permanent writer and no product consumer.
Authority theater would have been a switch on a consumer that does not exist,
reading a store that is not being written. So: make the graph live, measure it
from outside, and expose the one category it alone owns.
WRITE-PATH FAILURE SEMANTICS, TRACED FIRST. persist(:857) -> the authoritative
cacheSet snapshot:latest(:1404) -> the lineage gate(:1426) -> ledger(:1473).
Lineage failure cannot fail base retention (earlier, separate upsert), cannot
fail publication (the product write precedes the gate), cannot fail the Ledger
(lineageIndex defaults null; the ledger has its own guard), and cannot create a
new partial row (blankLineage() first, atomicity sweep after the catch). The
isolation this tranche needed already existed; only a mode was missing.
THREE MODES, ONE EVALUATOR. `lineageWriteMode` distinguishes OFF /
CANARY_LEASED / PERSISTENT_SHADOW, and the write gate and the status surface
both read it, so they cannot disagree. The canary is CONSULTED, never
converted: its <=4h absolute expiry, its dynamic evaluation and its
fail-closed parse are untouched.
AN AMBIGUOUS CONFIGURATION FAILS CLOSED. If a sport is named by both persistent
mode and an active lease, the two instructions disagree about WHEN WRITING
STOPS — the lease says 22:45, persistent says never. The dangerous reading is
the quiet one: an operator sets a bounded lease believing writing will stop
while persistent keeps it going. We cannot know which they meant, so that sport
writes nothing until the configuration says one thing. The sport allowlist is
the canary's own, so persistent mode can never widen past it.
THE OBSERVER MAY NOT ASK THE WRITER HOW IT DID. settleLedger returned
{settled:0,pending:0} — byte-identical to a healthy "nothing to settle" — while
1,444 rows sat unprocessed, and the watchdog believed it. So `lineageCoverage`
reads durable retained state only, and THE DENOMINATOR MAY NOT CONSULT
lineage_action: eligibility is "the row was published AND a natural key is
derivable from its own identity columns", neither of which the lineage path
writes. If expectation were derived from whether lineage exists, coverage would
be 100% by construction and the metric would be decoration. Zero-expected and
zero-written are kept as different answers.
THE LEGACY BOUNDARY IS OBSERVED, NOT DECLARED. `publication_id` is stamped only
by commitPublication, so the row itself says whether lineage ran. Verified on
production: 5,353 rows carry it — 4,234 complete actions plus exactly the 1,119
historical partial rows — and zero actions exist without one. No epoch constant
is invented; a date would have been a guess about when the writer was on.
publication_id NULL -> LEGACY_UNVERIFIED. Stamped but incomplete ->
LINEAGE_UNAVAILABLE, which is the honest answer for the 1,119 and is never
quietly rewritten as legacy.
THE GRADE-SHIFT BADGE IS UNTOUCHED. revised_from_grade answers "did the letter
change"; lineage answers "which published claim superseded which". Different
questions, and a test now fails if either route learns the word lineage.
Suite 397/5,496/0 · tsc 0 · 15/15 teeth.
TWO OF MY OWN TESTS WERE VACUOUS AND A TOOTH FOUND IT. Tooth 2 came back green
because the isolation tests asserted the slate was published — true whether or
not the exception propagated — while never reaching the lineage gate at all:
the fake grader never fired `onGraded`, so the collector stayed empty and
persistedRows stayed null. Fixed by firing the hook and counting the commit.
A green teeth run means the test is missing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CQJeAG8vcDoL5zkiaJyVb8
|