Past-cap census as a first-class calibration diagnostic

[MaxGhenis]

What

Deliverable A of #492: the past-cap census ships in calibration_diagnostics.json (and every artifact written through write_calibration_diagnostics, including the read-only scorer). Pure observability — no solver behavior change.

New past_cap_census block in diagnostics_payload (schema 4 → 5):

• initial_past_cap / final_past_cap — rows at or past target_loss_cap under the initial / final estimates
• escaped — past at initialization, back inside at final
• frozen — past at both ends
• pushed_out + pushed_out_rows — rows that started inside the cap and were pushed past it by the solve (name, init_rel, final_rel), worst final miss first

Misses are scaled with the loss's own rule max(|target|, 1), so "past cap" here means exactly "zero gradient in the objective". The cap is read from result options in both recorded shapes (calibrate's target_loss_scales.cap and score_targets' top-level target_loss_cap); results with no recorded cap census as null.

Why

The measured Build N census (#492): 929 rows past cap at init, 377 at final, and 17 rows pushed out — in-cap rows abandoned past the cap because the 101st percentage point of error is free (LA TANF 0.47 → 5.54; four of the six broken medical states got worse while past cap). Nothing reported which rows the solve wrote off. Now the release artifact does.

Scope

Diff is entirely inside packages/populace-calibrate (PR #491 owns the builder/populace-data files; no overlap). The builder and tools/score_us_fiscal_targets.py pick the census up for free through diagnostics_payload.

Testing

• Stub-result test pins the exact classification of every row class, the max(|target|,1) scale basis (a 0.5-valued target censuses on scale 1, not raw relative error), the ≥-boundary (a row exactly at the cap counts as past), and the worst-first ordering
• Integration tests through calibrate (frozen hopeless row), score_targets (score-shape cap pickup; no motion ⇒ escaped == pushed_out == 0), and calibrate_l0_refit (merged-options cap)
• Census identities asserted: initial_past_cap == escaped + frozen, final_past_cap == frozen + pushed_out
• Full populace-calibrate suite green locally; ruff format + ruff check clean

Refs #492.

🤖 Generated with Claude Code

[MaxGhenis]

Owner gate: found and fixed a cross-package schema break — this PR moved populace-calibrate to diagnostics schema 5 but populace-data's publish contract pinned 4 and hard-rejects mismatches (contract.py _check, 'this library publishes version 4'), so the next release build would have gone green and then publish-failed on schema version — the same builder-pass/publisher-reject class #491 just closed, reintroduced across packages.

Pushed the lockstep commit: contract + fixtures to 5, plus test_calibration_diagnostics_schema_lockstep in the builder suite (the one place that imports both packages) so the two same-named constants can never drift silently again. Suites green (data 133, calibrate diagnostics, lockstep). Sol confirm next, then merge.