feat(level-factors): prior pushforward + surfaced t2 psense warning (#389 finding 3)

[ethanbuckley]

Note

Drafted by an LLM-based AI tool (Claude Code/Opus 4.8).

Second tranche of #389, stacked on #445 (base feat/lrp389-level-factors-runplan). Addresses finding 3: the level family lacked the estimand-scale prior_pushforward.csv the ITT/gain families emit, and its key-findings headline stayed unqualified even when power-scaling flagged the t2 term for prior-data conflict (the warning was hidden in a collapsed prior section).

Prior pushforward (the well-defined half)

The ITT prior_pushforward assumes a tau/tau_i structure that the level model doesn't have — its t2 effect is b_grp_time[1] computed by the bespoke level_t2_marginal_effect (which nets the full per-timepoint group vector + group × ability term out and adds back only the t2 contrast at mean ability). So:

• level_t2_marginal_effect gains group="posterior" | "prior" — the same net-out transform now runs on either group (default "posterior", byte-identical);
• new level_prior_pushforward pushes the prior on the t2 contrast through that transform and returns the same items-scale schema as prior_pushforward, so config.json and any consumer treat both families' prior checks uniformly;
• fit_level_factors writes prior_pushforward.csv beside the posterior ROPE card.

The level model has a single randomised t2 term, so the pushforward is well-defined here — unlike the HSGP-curve / correlation-matrix families that stalled the broader #381 pushforward ask. lf-001 dev fit emits a sane check: prior on the t2 contrast median ≈ 0, ±11 words of 79 (wide but not absurd).

Surfaced t2 psense warning

_kf_build_level_factors now surfaces a caution bullet right after the headline when psense_summary.csv flags b_grp_time[1] (e.g. "potential prior-data conflict"), reading the diagnosis via a new _kf_psense_diagnosis helper. It's a graceful no-op when psense is absent or the term is clear, so fits that never ran power-scaling — and the synthetic golden test — are unchanged. With the warning the level key-findings box is exactly 5 sentences (headline / warning / confidence / rope / causal), within the cap.

Tests

• level_prior_pushforward reads the prior group (a +100-shifted posterior proves it isn't read) and returns the schema;
• the KF warning appears when the t2 term is flagged, is absent when the diagnosis is clear, and the golden level case (no psense) is unchanged.

ruff, npm run format:check, npm run spellcheck all clean.

Still a follow-up

Finding 3's remaining piece — a direct prior-grid sensitivity for the flagged outcomes (W, L, P, B, N) — is refit-gated (VM) and out of scope here. Findings 1 (t2 estimand decision), 2 (prior recentring) and 5 (SP/RW naming) remain as tracked in #445's description.

Assigning to a human for review and merge.

[Copilot]

Pull request overview

Adds level-factors parity with other estimand-reporting families by (1) emitting an estimand-scale prior_pushforward.csv for the t2 contrast and (2) surfacing a power-scaling (psense) caution directly in the level-factors key findings when the t2 term is flagged, rather than leaving it hidden in collapsed sections.

Changes:

• Extend level_t2_marginal_effect to run the same “net-out then add-back t2” transform on either posterior or prior draws, and add level_prior_pushforward to export the prior-on-estimand check in the standard schema.
• Write prior_pushforward.csv during fit_level_factors alongside the t2 ROPE outputs.
• Add key-findings logic + tests to surface a t2 psense warning immediately after the headline when psense_summary.csv flags b_grp_time[1].

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
src/language_reading_predictors/statistical_models/reporting.py	Adds prior-group support for the level t2 estimand transform, implements level_prior_pushforward, and surfaces psense-based warnings in level-factors key findings.
src/language_reading_predictors/statistical_models/pipeline.py	Persists prior_pushforward.csv for level-factors fits using the new reporting helper.
tests/statistical_models/test_reporting.py	Adds coverage that the level prior pushforward reads from the prior group and returns the expected items-scale schema.
tests/statistical_models/test_key_findings.py	Adds coverage for the new level-factors psense warning sentence (present when flagged, absent when clear/absent).