Ship the generator that produces ch26's pinned prediction artifacts

[stefan-jansen]

What this is

PR #419 took ch26 green with the notebook half of fix/ch26-pinned-prediction-fixture. The half that generates the artifacts those notebooks pin never landed. ch26 is green today only because the artifacts are already committed to the test-data repo — the next fixture regeneration drops them and the job goes red for a reason nobody will connect back to this.

This is that missing half.

Why it is a rebuild rather than a merge

The original branch is 14 commits behind main and merging it would have reverted PR #420: a 2,731-line deletion against current main, taking out the eoa seal gate, input_lineage, and the coverage bar.

Its notebook changes and its tests/overrides.yaml changes are already upstream via #419 (verified — both are byte-identical to main). So the entire remaining delta is one file, and it is purely additive: +406 lines, 0 deletions.

main's _copy_backtest_artifacts (from #420) and these four functions were a three-way conflict only because both sides inserted at the same point in the file. Both are kept — they share no symbols, and main() calls each.

What it adds

• ensure_pinned_predictions / preflight_pinned_predictions — give every pinned hash both a registry row and a materialized artifact, idempotently. Preflight decides every destination and validates the layout before any registry is recreated, so a caller that skips it fails the same way rather than failing after the destination has been mutated.
• _pinned_prediction_universe — one universe shared by every pinned artifact and by the fixture's own data. Reading it off whichever artifact sorted first gave 50 symbols meeting a 56-symbol fixture universe in two names, which would have left two assets per date and made every cross-sectional correlation ±1 by construction.
• _copy_rows_onto_existing_schema — projects onto the columns the destination actually has, so a fixture written before a column was added stays usable instead of failing on it.

Review finding, addressed in the second commit

roborev flagged that preflight compared (symbol, timestamp) only, so two streams could share every key and still carry different actual values for it — the notebooks would then score model error and realized return against different outcomes with every breadth and key check passing.

Not reachable with today's pinned set (the identity assertion pins all three us_equities_panel hashes to fwd_ret_1d, and they agree row for row against production, 0 nulls), but the pinned set is data and the next hash added need not share the label. The comparison now carries the outcome column, and null targets inside the shared universe are rejected.

Verified not vacuous: perturbing one artifact's actual by a constant is rejected; the key-only comparison it replaces accepted the same perturbation silently.

Verification

• ruff check / ruff format clean; all pre-commit gates pass.
• preflight_pinned_predictions('us_equities_panel', ...) runs green against the production registry.
• No notebook, no override and no fixture content changes — this is generator code only, so no CI job's current result should move.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.