vyndr/specs/MASTER-STATE-2026-08-17.md

VYNDR — MASTER STATE, 2026-08-17

This is the board. Every future order reads it first.

Derived from the repo and from live database checks on 2026-08-17, not from any summary. Every line is marked VERIFIED (checked this session, with the check shown), INFERRED (follows from something verified), or UNVERIFIED (could not be checked from here — do not act on it as fact).

Repo HEAD: 6c34af3 "checkpoint: chain shadow, WNBA possession feed, baseball chain" — VERIFIED pushed to gitea/main (git ls-remote gitea main → 6c34af34140b6dd…).

---

🔴 0. STOP — LIVE PRODUCTION REGRESSION, FOUND THIS SESSION

model_snapshots (the retention / measurement moat) stopped being written on or about 2026-08-15. VERIFIED by direct row counts:

game_date	model_snapshots	ledger_entries
2026-08-12	14,718	2,187
2026-08-13	12,004	1,368
2026-08-14	7,080	1,834
2026-08-15	378	1,988
2026-08-16	0	2,020
2026-08-17	0	220

The rest of the pipeline is healthy: ledger_entries is still writing, and closing_captures newest captured_at = 2026-08-16T23:01:49Z. So this is not a dead cron — it is one write path failing.

Cause — VERIFIED by code + schema check

1. src/services/retentionService.js:165 declares chain_shadow: null on every retention row (added by the chain-v1 order so a PostgREST bulk insert has one consistent shape).
2. Migration 038_chain_shadow.sql was never applied — VERIFIED: select chain_shadow from model_snapshots → "column model_snapshots.chain_shadow does not exist".
3. retentionService.persist() catches the error into out.error; the caller logs a warning and the snapshot continues, because retention is best-effort by design. So every insert fails and nothing pages.

That is exactly the failure the chain-v1 report flagged as a blocking precondition, verbatim: "If the column does not exist, every retention insert fails, and retention is best-effort, so it fails SILENTLY — the exact 'retention writes nothing' shape the S64 ops alarm exists to catch." The commit was pushed and deployed; the migration was not applied.

Consequence — the accrual clock is FROZEN

A8 (§5) and the chain-vs-counter head-to-head both accrue on model_snapshots. Neither has gained a row since 2026-08-15.

The fix is one of two things, and NEITHER was done here

This order is documentation-only, so nothing was changed. The options:

• (a) Apply migration 038. Correct and complete. Needs DB DDL access, which is unavailable from this environment (§9).
• (b) Revert retentionService.js:165 (chain_shadow: null). Restores retention immediately with no schema change; the chain shadow then has nowhere to land, which is acceptable because it is served by nothing.

(b) is the faster unblock and requires no credentials. This is the single highest-priority action on the board.

---

1. STORAGE — and why WNBA forces the Pro decision

item	state
closing_captures	1,294,740 rows — VERIFIED. Down from the ~4.2M peak.
closing_captures where missed_reason='missed_window'	0 — VERIFIED. Fix B2 held; the bleed is stopped AND the historical rows are gone.
lock_lines	ABSENT — VERIFIED ("Could not find the table"). Dropped, writer retired.
ledger_entries	30,209 rows — VERIFIED
model_snapshots	180,410 rows — VERIFIED
statcast_history	19,798 rows — VERIFIED (point-in-time window accruing)
team_defense	403 rows — VERIFIED
Total DB size vs 500 MB cap	UNVERIFIED. pg_database_size needs DDL/SQL access, unavailable from here. Last recorded figure was 406 MB (81%) after the lock_lines drop. Given closing_captures fell from ~4.2M to 1.29M rows since, the true current figure is probably materially lower — but that is INFERRED, not measured. Re-measure before any capacity decision.

WNBA footprint — MEASURED, and it contradicts the number in the order

The order states WNBA makes Pro a requirement at ~475 MB/season. Measured value is 42.6 MB/season — an order of magnitude lower.

392 bytes/event   (typed columns + 24B tuple header + ~40B PK index entry)
× 411.9 events/game × 264 games  =  42.6 MB per season   (0.158 MB per game)

Method: 16,114 real events pulled over 2026-08-01..14 through the shipped parser; JSON-serialised at 596 B/event; typed-column factor 0.55 plus Postgres row and index overhead. VERIFIED (scripts/wnba-possession-verify.js).

The Pro conclusion may still be right, but not from this number. WNBA v1+v2 together ≈ 58 MB. Against the last-known 94 MB headroom that is ~62% of what is left — enough to block a second sport and next season, not enough to be a crisis. Where 475 MB came from is unknown; do not plan against it.

---

2. MLB REPAIR — VERIFIED present in the repo

item	evidence
Read integrity	scripts/read-integrity.js REGISTRY = 34 readers (VERIFIED by count). safePaginate/pageSafe used in 26 files; 16 scripts export named READS.
Version stamp alignment (A3)	src/config/modelVersion.js is the single declaration. Live rows carry engine1@2026-08-07-fullwindow — VERIFIED from model_snapshots.model_version.
Settlement	ledger_entries settled = 27,941; model_snapshots settled = 54,736. Backlog of unsettled rows with game_date < 2026-08-16 = 0 — VERIFIED. Settlement is healthy.
As-of context (A4)	hitsFactorContext.build(sb,{asOf}) present.
Factor-input freeze (A5)	model_snapshots.factor_inputs column PRESENT, 4,527 rows populated — VERIFIED. A5 is deployed and was accruing.
Join keys / shadow fire (A6/A7)	factor_inputs.would_fire present on 1,445 of 3,389 rows sampled 08-05..08-14 — VERIFIED. The factors that never fired are firing in shadow.
dCLV (B1)	ledger_entries.dclv_state populated — VERIFIED.
Prod deploy sha	UNVERIFIED. vyndr.app/api/health → 404, www.vyndr.app → 503. Cannot read the container's sha. INFERRED to be at or near 6c34af3 because (i) it is on gitea/main and (ii) the retention failure in §0 requires that commit's code.

---

3. THE CHAIN ENGINE — built, shadow, served by nothing

Files VERIFIED present: chain.js (325L), baseballChain.js (272L), chainShadow.js (359L).

• chainFn slot added — was described in the header and absent from the code for the module's whole life. Defaults to identity-on-p, so prior callers are byte-identical.
• Sign-aware correlation — clamp [0,1] → [-1,1], interpolating toward the Fréchet bound the sign selects. The positive branch is arithmetically unchanged. Basketball's negative usage-competition case was previously inexpressible, not merely mismodelled.
• redistribute on both readings + exported prepareAtoms, so a redistribution cannot reach one reading and make selfCheck flag an inconsistency the model just manufactured.
• MLB chainFn fires via skillProjection.paOutcome → Binomial(PA) — 99.6% of the board (9,752/9,792), sole refusal reason no_batter_profile.
• The PA→game conversion was never broken. paDistribution is a mean-preserving two-point mixture and atLeast(Binomial(n,p),1) is 1−(1−p)^n averaged over n. The real defect was the input E[PA] — the shadow passed no lineup slot, so every hitter ran on DEFAULT_PA = 4.1. Fixed; posted slot now reaches 88.8% of props.

DIAGNOSTIC RESULT — state it precisely

Quintiles of opposing-pitcher K%, over side, n=3,705:

Q1 (soft) +0.0749 → Q5 (hard) +0.0256   spread +0.0494   Pearson r = −0.119
below-counter share monotone across all five: 27.8 → 30.8 → 34.1 → 35.0 → 38.6

The result is BOTH, and the order's phrasing ("signal") is half of it:

1. A difficulty-correlated component (~+0.049) — the chain conditions on the matchup.
2. A uniform +0.026 offset surviving into the hardest quintile — it does not move with the matchup, so it is not conditioning. Unexplained. Calling the whole result "signal" buries it.

And the correlation is near-mechanical: the chain reads opposing-pitcher K% directly and the counter reads nothing about the pitcher, so a monotone relationship is close to proof that the wiring reaches the number — not that the adjustment is correct. Only settled outcomes can say that.

CALIBRATION_DEPLOYED = Object.freeze([]) — VERIFIED at snapshotService.js:301. The counter still serves. The grade was frozen throughout, proven byte-identical with the shadow on and off.

---

4. WNBA — CORRECTION: built, NOT ingested

The order says the possession feed is "ingested". It is not. Nothing WNBA exists in the database. VERIFIED:

wnba_player_game        → "Could not find the table"   (migration 039 unapplied)
wnba_possession_event   → "Could not find the table"   (migration 040 unapplied)
wnba_season_context     → "Could not find the table"   (migration 041 unapplied)

Three further premise corrections, each VERIFIED:

• NOT nba_api playbyplayv3. The package is not installed (ModuleNotFoundError; pinned only in the offline python service), and stats.wnba.com is WAF-blocked — HTTP 000 on every call after the first; cdn.wnba.com returns 403 from failover-waf.nba.com. The source is ESPN, with wehoop parquet as the bulk twin.
• scoremargin is not a served field anywhere. It is derived from homeScore − awayScore (exact, but derived).
• Basketball-Reference cannot serve as-of-date aggregates — season-to-date as of today, no date parameter. It is captured dated, forward-only, and its real value is an independent second opinion on usage.
• wehoop has no lineup/on-off data (14 data dirs, none of them lineup), so it was deliberately not given a validation table — that would be a second ~30 MB copy of the same stream with no reader.

What IS built and measured (in memory, over the live season)

espnWnbaAdapter (377L), wnbaUsageService (215L), wnbaPbpAdapter (253L), wnbaLineupState (197L), wnbaSeasonContextService (219L) — all VERIFIED present.

16,114 events · 40 games · 402.9/game · 0 fetch errors  (08-01..08-14)
score_margin / game_seconds_remaining / period   100% each
substitutions 2,175 — BOTH players resolved on 2,175 of 2,175
lineup fold: 40/40 games resolve to a COMPLETE 5-a-side
states: competitive 14,177 · clutch 864 · lopsided 840 · decided 233
as-of leakage: NONE at three cutoffs

⚠️ CORRECTION on "players-on-court NULL"

The order lists players-on-court as NULL with lineup reconstruction pending. Lineup reconstruction is BUILT and works — wnbaLineupState.foldLineups folds the on-court five forward from substitution events and resolves 40/40 games to a complete 5-a-side. What is true is that no column stores it, because no table exists. It is computed, not persisted.

Dormant / honest gaps that remain:

• Team-level only for redistribution attribution today — per-player on/off needs the fold joined to outcomes, which needs the table.
• The feedback layer is capable but unproven: no chainFn consumes it yet.
• A real parse bug was found and fixed: ESPN drops the colon under a minute ("56.2", "5.7"), so a colon-only clock parser discarded the final minute of every period — the clutch/garbage-time window. Coverage 86.6% → 100%.

NEXT for WNBA = the chainFn (v3). Not started.

---

5. PENDING VERDICTS ON THE ACCRUAL CLOCK

A8 — do the MLB hits factors improve the forecast?

Spec: specs/a8-shadow-factor-gate.md, pre-registered before any accrual existed. Binding constraint is MIN_CLUSTERS = 40 distinct settled games, not MIN_N.

MEASURED 2026-08-17 (sampled 08-05..08-14, 1,000 rows/date cap):

rows with factor_inputs   3,389
  ...with would_fire      1,445
  ...AND settled          623   COMPLETE PAIRS   (MIN_N 500 → MET)
distinct settled games     20                    (MIN_CLUSTERS 40 → 50%)

Status: HALFWAY, AND STALLED. The row bar is met; the cluster bar is at 20/40 — and no new pairs have accrued since 2026-08-15 because of §0. The clock does not restart until retention writes again.

The order's "~Aug 17" date is not met and was never going to be met on clusters; it was a row-count estimate.

Chain vs counter — Brier on the (chain_p, counter_p, outcome) triples

Not accruing at all. The triple lands in model_snapshots.chain_shadow, which does not exist (migration 038). Zero rows. Blocked on the same fix.

Both verdicts are pre-registered, both wait on settled games, and both are currently gated on a manual "did we hit N" check — which is exactly what build-process upgrade (c) below removes.

---

6. THE REMAINING BUILD WAVE (ordered)

0. 🔴 Restore retention (§0) — everything downstream is blocked on it.
1. WNBA chainFn (v3) — usage × possessions × efficiency, contested opportunity, team-level feedback.
2. Lineup reconstruction → player-level redistribution — the fold exists; this is joining it to outcomes and persisting.
3. WNBA shadow — same fence as MLB: served by nothing.
4. Gate adjudication — A8 + chain-vs-counter through factorGate and the cumulative Bonferroni denominator.
5. NBA — content swap onto the same machinery; season opens October.
6. Soccer.
7. Media / design layer — F5 article media, E16/F8 cards, in-season hub: the surfaces that reflect all sports + correlations.

---

7. BUILD-PROCESS UPGRADES IDENTIFIED

• (a) NEXT — independent verifier on fresh context for the grade-freeze / 0-leaks invariant. The invariant is currently asserted by tests written by the same pass that wrote the code. §0 sharpens this: the failure that actually bit was not a leak but a silently-failing best-effort write with an unapplied migration. The verifier should check deploy preconditions (does every column the code writes exist?), not only the leak invariant.
• (b) SOON — deterministic gates from the harness. Promote read-integrity anchors to gates.
• (c) SOON — scheduled loop for the accrual verdicts, removing the manual "did we hit N dates" clock. §5 shows the cost of the manual clock: the accrual stopped two days ago and nothing noticed.
• (d) NAMED-FUTURE — data-dependency-graph audit of the codebase: find false sequential dependencies and missing real edges.

Add (e), earned this session: a migration/deploy interlock. A commit whose code writes a column that no applied migration provides must not reach prod. §0 is the second time a documented blocking precondition was flagged and then bypassed.

---

8. STANDING DOCTRINE + INVARIANTS

1. The served grade never moves without a human. Every model change ships behind a shadow first; CALIBRATION_DEPLOYED = [] and the counter serves.
2. Moat tables are never touched by an automated loop — ledger_entries, model_snapshots. (Note the irony of §0: the moat was not touched, it was starved. "Do not write" is not the only failure mode; "silently stop writing" is the other.)
3. Each sport is its own model. Machinery shared, content re-derived, factors start at α = 0 and earn their place through the gate.
4. Truth Law — no fabricated data anywhere; honest-absent beats invented; absence is not zero; label limitations in-band; provisional stays provisional until re-run; verify by inducing the real code path.
5. Nothing is proven. proven-status.js → the PROVEN set is EMPTY. validatedSkills() = {} for every archetype. The only proven feature is the incumbent counter itself.
6. Cumulative Bonferroni across the programme lifetime; α only ever shrinks; re-tests do not inflate it.

---

9. THE STANDING ENVIRONMENTAL BLOCKER

Four migrations are written and unapplied: 038, 039, 040, 041 — VERIFIED, all four target tables/columns confirmed absent.

DDL cannot be run from this environment:

• SUPABASE_DB_PASSWORD fails pooler auth (aws-1-us-east-1… → "password authentication failed").
• db.<ref>.supabase.co resolves IPv6-only → ENETUNREACH from WSL2.
• PostgREST reads and writes work fine — DDL is the blocker, not writes.
• Local .env has the transposed project ref (zmdnczhtdxcddszxttub; real is zmdnczhtdxcddsxzttub), so scripts need an explicit SUPABASE_URL= override.

Related IPv6 lesson: stats.wnba.com also fails over IPv6 and works over IPv4 (curl -4). Anything reporting a host unreachable from this box should be re-tested with -4 before it is believed.

---

10. TWO ANOMALIES WORTH A LOOK (neither investigated)

• 28 ledger_entries rows have game_date in the future (> 2026-08-17), newest 2026-10-20 — VERIFIED count. Could be futures markets or mis-dated rows. Unexamined.
• model_snapshots.captured_at returned NULL on an ordered limit 1 read. Unexamined; may be an ordering artifact or a genuine null.

---

Board compiled 2026-08-17 from repo HEAD 6c34af3 and live database checks. No code, schema, served path or accrual was changed by the order that produced this document.