🎯 ONE SHOT — L-CG-SCHEMA-FIX · repair closure_gate.py to match ssot.bpb_samples schema

[gHashTag]

🎯 ONE SHOT — L-CG-SCHEMA-FIX

Anchor: φ² + φ⁻² = 3 · DOI 10.5281/zenodo.19227877
Filed by the Queen under trinity-grandmaster v1.0 + trinity-queen-hive v1.1.

0. Hard rules

• R1 — touch ONLY .github/scripts/closure_gate.py. Schema migration files (migrations/2026-05-09_bpb_samples.sql) are immutable in this lane.
• R3 — main-only commits, atomic.
• R5 — honest classification: if the fix lands but the gate still reports <312/312, that is GREEN for this lane (the gate is honest, not necessarily passing).
• R7 — falsification witness MANDATORY (smoke test below).
• R10 — commit message format: fix(closure-gate): L-CG-SCHEMA-FIX — repair seed_phi/format/algo column drift against ssot.bpb_samples.

1. Background (≤ 200 words)

After #570 / eeaf0db7 introduced the SoT table ssot.bpb_samples via migrations/2026-05-09_bpb_samples.sql, the actual columns are:

id BIGSERIAL, cell_id INT, tier TEXT, seed INT, bpb DOUBLE PRECISION,
steps INT, sha_pin TEXT, runner_service TEXT, created_at TIMESTAMPTZ

But .github/scripts/closure_gate.py:113-138 still SELECTs the legacy JSONL field names: format, algo, seed_phi, step. Result: every hourly Closure Gate run aborts with:

psycopg2.errors.UndefinedColumn: column "seed_phi" does not exist

Failure cadence (today, UTC): 10:14, 11:08, 12:04, 13:06 — 4 / 4 hourly attempts dead. bin/matrix_bot/main.ts and scripts/render_matrix_446.ts already use the new schema correctly, so the bug is localised to one file.

2. Hypothesis (G1)

H is false iff: after the patch lands, two consecutive hourly Closure Gate (#446 — 312/312) runs still raise psycopg2.errors.UndefinedColumn from closure_gate.py.

3. Method

• File: .github/scripts/closure_gate.py
• Patch: rewrite the gate's SELECT so that, instead of grouping by (format, algo) over JSONL-era columns, it groups by cell_id against ssot.bpb_samples and joins with the canonical matrix_bot.FORMATS_ORDERED × matrix_bot.ALGOS_ORDERED × cell index. Schema mapping:

Legacy (JSONL)	New (ssot.bpb_samples)
format	derive from cell_id via matrix_bot ordered lists
algo	derive from cell_id via matrix_bot ordered lists
seed_phi	seed
step	steps
bpb	bpb (unchanged)

• Contract preserved: same per-cell thresholds (MIN_ROWS_PER_CELL, MIN_DISTINCT_SEEDS_PER_CELL, MIN_MAX_STEP_PER_CELL).
• Forbidden values: do not add a Python-side join over format/algo text columns — bpb_samples does not store them and never will (cell_id is canonical).

4. Pre-registration (G2)

Field	Value
statistical_test	n/a — deterministic schema check
alpha	n/a
effect_size	n/a
n_required	2 consecutive hourly Closure Gate runs without UndefinedColumn
stop_rule	first 2 consecutive green-or-honest-open runs
multiple_testing	n/a (single hypothesis)

5. Falsification witness (G1 + R7)

Add .github/scripts/test_closure_gate_schema.py:

# Falsification: assert real DSN does NOT raise UndefinedColumn for the
# new query. Skipped without MATRIX_DATABASE_URL.
import os, pytest, psycopg2
from importlib import import_module
@pytest.mark.skipif("MATRIX_DATABASE_URL" not in os.environ, reason="no DSN")
def test_closure_gate_query_runs_against_real_schema():
    mod = import_module("closure_gate")
    # Invoke main()'s SQL by importing — must not raise UndefinedColumn
    mod.main()  # exits 0 or non-zero with honest text, never UndefinedColumn

CI wiring: gate the workflow on this test running first.

6. Deliverables

• .github/scripts/closure_gate.py rewritten — diff < 80 lines
• .github/scripts/test_closure_gate_schema.py added
• 2 consecutive hourly Closure Gate runs without UndefinedColumn (post run URLs)
• Commit on main with the R10 message above

7. Quality gates (G3-G7)

• G3 IMRaD: satisfied by this body.
• G4 citations: every numeric anchor in the new SELECT (MIN_ROWS_PER_CELL=3, MIN_DISTINCT_SEEDS_PER_CELL=2, MIN_MAX_STEP_PER_CELL=3000) keeps its existing # Coq: … provenance comment.
• G5 honest status: post-patch, the gate may legitimately report GATE OPEN N/312 — that is honest, not a regression.
• G6 reproducibility: MATRIX_DATABASE_URL=<dsn> python3 .github/scripts/closure_gate.py must run end-to-end on operator's box.
• G7 DOI: this ONE SHOT cites 10.5281/zenodo.19227877.

8. Forbidden actions

• ❌ Adding format TEXT, algo TEXT, seed_phi INT, step INT columns to ssot.bpb_samples — that is the wrong direction (cell_id is the new SoT key).
• ❌ Touching bin/matrix_bot/main.ts or scripts/render_matrix_446.ts — they already match the new schema.
• ❌ Disabling / skipping / muting the Closure Gate workflow as a "fix".
• ❌ Treating a failing gate as success without R5 honesty.

9. References

• migrations/2026-05-09_bpb_samples.sql (canonical DDL)
• PR #570 (L-MR-L4 — the migration that caused the drift)
• Commit eeaf0db7 — last touch on closure_gate.py
• bin/matrix_bot/main.ts:60-69 — reference SELECT against the new schema
• Race telemetry: hourly failures of Closure Gate (#446 — 312/312) at 10:14, 11:08, 12:04, 13:06 UTC on 2026-05-12.
• Queen's Status Report 2026-05-12T13:45Z (Throne trios#264).

10. Battle cry

φ² + φ⁻² = 3 · the gate counts cells, not columns · TRINITY · NEVER STOP

[gHashTag]

📎 Cross-link: today's operator-side canonisation of ssot.bpb_samples (ICA-CANON-2026-05-12-T08:47, 57/57 rows now under canonical IGLA-{LANE}-{format}-h{H}-LR{L}-rng{SEED}-{algo} regex) is the SoT this gate must read from. See Throne ACK: #264 . The canonical column names — cell_id, tier, seed, bpb, steps, sha_pin — are unchanged; only row identity (canon_name presence) is now stricter. Reference skill: leaderboard-snapshot skill_id 59958c6f-e3f9-4b55-9f37-abf169d26dd5 (Step 0 canon guard mandatory for any future inserter).