project_wnba_lineups_gap

Original note (2026-06-18) below — "NOT committed" is stale:

BUILT 2026-06-18 (not deployed, not run e2e): RotoWire projected-minutes ingest in SDS internal/rotowire/ — shared authenticated league-aware client (URL /basketball/ vs /wnba/)

• per-league refresher (enumerates today's teams from {schema}.games, fetches RotoWire, name-matches to {schema}.players.player_id, upserts {schema}.game_lineups with projected_minutes
• [internal detail removed]

WNBA pick projections are injury-aware (official injury PDF → Out/Doubtful/Questionable lockdown gate works) but NOT starting-lineup or projected-minutes aware. Verified 2026-06-18. The infra exists; it's a bug, not a missing feature.

Two distinct gaps:

1. SDS WNBA lineups ingest is empty. SDS owns lineups (route /{league}/lineups/{id}, internal/adapters/wnba/lineups_refresh.go, table wnba.game_lineups). The route returns 200 but players:[] for every game incl. finals. Refresher pulls stats.wnba.com/js/data/leaders/00_daily_lineups_{date}.json; its own comment admits WNBA daily-lineups coverage is unreliable and it returns clean on 404. So data-hydrator's assembleMultiSport deliberately bypasses SDS /lineups and uses a roster-fallback (assembler.go:1412 comment: "WNBA lineups-refresh ingest is broken upstream… separate fix"; roster-fallback at ~1564, LineupType=PROJECTED, projected_minutes=0).

2. RotoWire (projected minutes) lives ONLY in lineup-service — not SDS. assembleNBA calls lineup-service /lineups/game/{slug} → real lineup + RotoWire minutes; assembleMultiSport (WNBA) never calls lineup-service. lineup-service is NBA-only (getGamesWithLineups queries nba_game_id, NewNBALineupFetcher), so RotoWire GetTeamMinutes is never called for WNBA teams (the client itself is league-agnostic — /basketball/...?team=ABBR — and could serve WNBA). SDS has zero RotoWire code. game_lineups has a projected_minutes column but nothing fills it from RotoWire.

Plan: SDS is SoT for lineups; lineup-service is slated to retire (see project_sds_is_the_data_layer, project_sds_cutover_plan). Do NOT extend lineup-service. Recommended fix: port the ~250-LOC RotoWire client into SDS's wnba adapter, parameterize the URL by league (/basketball/ vs /wnba/), fetch per WNBA team, upsert starters + projected_minutes into wnba.game_lineups, wire into the lineups refresh cron; add ROTOWIRE_USERNAME/PASSWORD + PROXY_* to SDS env; then point data-hydrator WNBA at SDS /lineups (currently roster-fallback). This solves WNBA lineups+minutes WITHOUT the big NBA-only phase-5e projector (project_wnba_snapshot_depth_gap). NBA deprecation of lineup-service is the separate phase-5e lift.

ROOT-CAUSED 2026-08-24 (Maya Caldwell case) — THREE stacked failures

User bet OVER 4.5 pts on Caldwell (1630471, MIN) off a RotoWire starter listing; the board served UNDER. She played 25.4 min / 9 pts. Over won.

1. The lineup arrives AFTER TIP — WRONG, CORRECTED BELOW. See "CORRECTION". GET SDS /wnba/lineups/{game_id} DOES hold source: "RotoWire Confirmed Lineups" with Caldwell role: starter, projected_minutes: 26. But source_observed_at = 2026-08-25T01:54Z against a 00:00Z tip — +114 min. Checked the 14 most recent games: 0 pre-tip, 14 after tip, all at a suspiciously fixed +114/+144 min. The confirmed-lineup cron runs every minute (wnbaConfirmedLineupsCron, cmd/worker/main.go:362) so the poll is not the problem — we appear to only capture RotoWire's page once it flips post-game. The authenticated PROJECTED path is gated on jobs.rotowireClient != nil (main.go:369) i.e. ROTOWIRE_USERNAME/PASSWORD.

2. The date-wide route is 501 for WNBA. GET /wnba/lineups/date/{date} → {"error":"capability not implemented for this league"}. Only NBA implements LineupExtAdapter.FetchLineupsForDate (adapters/nba/lineups.go:335); WNBA's adapter has only per-game FetchLineup (adapters/wnba/lineups.go:156). projections-engine ingest/lineups.py calls exactly that 501 route — and its only caller repo-wide is the one-off scripts/lineup_game4.py. Dead code pointed at a dead route.

3. Even with the data, nothing would consume it. serve/minutes_projection.py builds minutes purely from the last 10 boxscores (RECENT_GAMES_WINDOW=10). simulate/minutes_conservation.py:16-18 admits in its own docstring: "no lineup-confirmation feed reaches this module" — it infers starter/bench from predicted q50 RANK, so a promoted bench player is still classed bench. models/training_set.py:436-437 justifies the leak-prone is_starter feature "because at inference we now have the lineup endpoint" — false in the shipped path; slate_scorer.build_rotation_snapshot takes the last historical row and never overrides it. Injury status can only REMOVE a player, never promote one. Also slate_scorer.py:178 drops anyone whose last game was <12 min (min_recent_minutes=12.0) — a promoted deep-bench player can vanish from the board entirely.

The system flagged its own blindness and shipped anyway. DH prop_diagnostic carried role_trend:"lineup_promotion" (run 1) AND minutes_source_confidence:"diverge_low_confidence" on the very pick that lost. Nothing downstream consumes those flags to veto or re-price.

⚠️ The served board contradicted its own math: nightly card ranked her UNDER 4.5 #1 in the game, grade B, while carrying value_flag:false, value_edge_pct:−1.04, kelly_stake_units:0. grade_v2 is calibrated P(hit)/ juice, NOT edge — the known trap, live on the card.

SDS already computes the correction it ignored: her starter split reports +11.3 min (t=7.23) and +4.0 pts (t=3.97), both verdict measured, driver minutes. 21/32 prior starts (65.6%) cleared 4.5 vs 35% over her bench-heavy L20. Same player, opposite side of the line, on one binary feature.

Market agreed with the user: under 4.5 moved −130 → −118 (CLV −2.4c), under 5.5 −164 → −138 (−4.1c); grader tagged both bad_price.

---

CORRECTION 2026-08-25 — the data WAS there, 16 hours early

The "+114 min after tip, 14/14 games" claim above is WRONG. It read source_observed_at off GET /wnba/lineups/{game_id}, which returns only the LATEST snapshot — post-game writes bump that field. It does not show when we first knew.

wnba.lineup_status_history is the real record. For Caldwell (1630471), game 1022600282, tip 2026-08-25T00:00Z:

08-24 07:24Z  -        -> PROJECTED  RotoWire Expected Lineups    (T-16h36m)
08-24 23:37Z  PROJECTED-> STARTING   RotoWire Confirmed Lineups   (T-23m)

ROTOWIRE_USERNAME/PASSWORD ARE SET on the SDS worker and the authenticated job works. Verified live: wnba.rotowire_lineups at 07:24Z wrote projected_teams=6 projected_starters=30 ~16h before tip. Do NOT re-diagnose this as a credentials problem.

Cheapest fix is on the PE side, not SDS. projections/ingest/lineups.py calls GET /{league}/lineups/date/{date} → 501 for WNBA (only NBA implements LineupExtAdapter). But GET /wnba/lineups/{game_id} works today and returns role + projected_minutes per player. PE builds slates per-game and already has game_id, so pointing it at the per-game route avoids writing any Go at all. (Implementing FetchLineupsForDate for WNBA is the tidier fix and matches PE's existing dead code.)

⚠️ SEPARATE, UNRELATED BREAKAGE: runWNBALineupsRefresh fetches https://stats.wnba.com/js/data/leaders/00_daily_lineups_YYYYMMDD.json, which now 302-redirects to HTML → parse: invalid character '<'. It errors on every run (cron 12,27,42,57, i.e. every 15 min) and has been filling the worker log. That feed is dead upstream; RotoWire is the live source. Either retire the job or repoint it.

⚠️ The projected job also rejects teams with expected exactly five unique starters; received 0/6/7 — some RotoWire pages parse to the wrong count, so projected coverage is intermittent even when it runs.

---

ALL FIVE FIXES SHIPPED 2026-08-25

Verified live: GET /wnba/lineups/date/2026-08-24 → 200, Caldwell role:starter, projected_minutes:26. Nightly board now serves valued_count / no_value_count.

⚠️ SCOPE — 2A moves MINUTES, not props. is_starter is a feature of minutes_v2 ONLY; it is in none of the nine per_stat.py STAT_FEATURES presets, and slate_scorer predicts stat quantiles at :410 BEFORE minutes at :425 without threading predicted_minutes. Making a promotion reach the POINTS line needs that reorder — a numeric change to every prop, needing its own measurement. Tonight's bad under is stopped by FIX 3, not FIX 2A.

⚠️ THE BOARD IS ESSENTIALLY ALWAYS NEGATIVE-EV. Measured 5 consecutive WNBA slates (08-20..24): core tier value_flag:false on 18/18, 30/32, 21/21, 28/33, 21/24. A hard EV floor would EMPTY the core card on two of five nights. So NIGHTLY_REQUIRE_VALUE exists but defaults OFF; the fix shipped is demotion + census, not exclusion. This is the open question, and it is bigger than the Caldwell pick: either p_hit is miscalibrated (memory: grade_v2 ≈ market) or we should not be serving a core card most nights.

⚠️ bestEdge → bestPick returns (valued, edge). A single float CANNOT express "any certified pick beats every refused one" — collapsing it to a float is how the first attempt silently did nothing. Test pins the inversion.

⚠️ I destroyed internal/adapters/wnba/status_history_test.go (untracked WIP) by mv-ing two same-named files into one temp dir. It referenced injuryStatusHistoryQuery, which does not exist in the wnba package — it never compiled, which is why go test ./internal/adapters/wnba/ was already failing. The NBA sibling survives at internal/adapters/nba/status_history_test.go and internal/handlers/status_history_test.go exists. Use git stash push -u, never mv.

CORRECTION 2026-08-25b — the NBA daily-lineups feed is NOT dead

An earlier recon claimed "the identical NBA feed is dead the same way" and I repeated it. Wrong — do not retire the NBA cron. Measured directly:

stats.nba.com  .../00_daily_lineups_YYYYMMDD.json  -> 200 application/json
   20260115 games=9   20260210 games=4   20251220 games=10   today games=0 (offseason)
stats.wnba.com .../00_daily_lineups_YYYYMMDD.json  -> 200 text/html 114KB error page

The NBA job logs nba lineups refresh ok ... games=0 errors=0 because it is the offseason, not because it is broken. Only the WNBA host redirects to an error page. cmd/worker/main.go NBA cron */15 * * * * stays.