deficienta: publish episodes, fix the strazi asymmetry, enforce the contract

[tiXor-code]

CMTEB publishes three states — Oprire, Deficiență (pressure/temperature below spec) and normal. The headline days counts only oprire+ACC, which is correct and deliberate. But deficiency was being flattened to a bare day count, and on streets not even that.

Measured on the current published bundle: 15,747 deficiency PT-days in 2025 against a 26,559-day headline (59%), rising to 73% year-to-date in 2026, with ~85% falling on days no outage touched.

These three must land together

validate's new key constants require the fields publish now emits. Landing validate alone would FAIL the next nightly on all ~11,700 entity-years and block the release.

1. Publish deficiency episodes (PT only — streets have no episode array)

The display record is hoisted above the severity branch and shared, so oprire and deficiency cannot drift apart. Three additive keys: episodes_deficienta, episodes_count_deficienta, est_hours_deficienta.

Strictly additive. The deficiency path still continues before pt_days, city_eps, pt_epn and pt_hours. test_deficienta_episodes_published_without_touching_headline proves it end-to-end through publish.build(): with one oprire and one deficiency episode, days == 2, episodes_count == 1 and est_hours == 24.0 while days_deficienta == 3 and the deficiency counterparts fill.

2. Emit days_deficienta on the strazi ndjson year object

defi was already computed there as the emission gate and thrown away. So street-years shipped deficiency runs with no counter to reconcile them against — while ARTIFACTS.md:91 claims the union invariant holds "for every PT-year and street-year". Verified against the live bundle: absent on all 6,670 street-years.

3. Enforce the contract in validate.py

It mentioned deficiency exactly once, in a docstring, and asserted nothing.

check	level	why
runs_days_union	FAIL	the ARTIFACTS.md:90-91 guarantee, previously unenforced
ndjson_year_keys	FAIL	a missing key renders NaN on the site
deficienta_reconciliation	FAIL	exact biconditionals by construction
rankings_row_keys	FAIL	exact contract key set per row
deficienta_non_vacuous	WARN	heuristic, not an identity

deficienta_non_vacuous FAILs only if every namespace-year is zero. A single quiet year is plausible upstream, and blocking the nightly release on a CMTEB editorial change would invert the design intent.

Two subtleties worth reviewing:

• The union check unions day-of-year numbers, it does not sum run lengths. Classes legitimately overlap (a day can be both avarie and programat — ARTIFACTS.md says so explicitly), and summing would FAIL on correct data.
• The ranking key check is a pre-pass, before the existing breaks. Those stop at the first bad row, so a key regression further down would never be seen. It also fixes a latent crash: a malformed row previously raised KeyError outside any try, killing validate with a traceback instead of a FAIL.

Verified against real data before enforcing

Ran the new invariant over the live published bundle:

pt: 5054 entity-years | headline union mismatches=0 | deficienta mismatches=0
st: 6670 entity-years | headline union mismatches=0 | days_deficienta ABSENT on 6670

The invariant is genuinely true in production, so these checks will not produce false FAILs. And end-to-end (new publish → new validate) reports 0 FAIL with all five new checks PASS.

Notes

• Detail strings carry only slugs, years and integers, never cause_raw, so untrusted CMTEB text cannot forge PASS/FAIL lines in the CI log (SEC049).
• tests/test_publish_shapes.py now imports the key sets from validate instead of redeclaring them, so the CI-visible check and the db-gated check cannot drift apart.
• Its fixtures were also made non-vacuous: strazi-2025.json was [] and both ndjson year objects were {"days": 3}, so several of these checks would have passed while asserting nothing.

45 passed, 15 skipped.

🤖 Generated with Claude Code

[claude]

Reviewed the full diff against CLAUDE.md's contract rules. Verified: the deficiență aggregation in publish.py is genuinely additive — epd is hoisted but the deficiență branch still continues before touching pt_days/city_eps/pt_epn/pt_hours, and the new parallel pt_eps_defi/pt_epn_defi/pt_hours_defi maps never feed the headline structures. The strazi days_deficienta fix correctly reuses the already-computed defi gate variable (pipeline/publish.py:798). validate.py's new checks are sound: _run_day_set unions day-of-year numbers rather than summing run lengths (correctly handles legitimate avarie/programat overlap), the rankings key pre-pass runs before the existing breaks so it also fixes a latent unguarded KeyError crash, and deficienta_non_vacuous is correctly scoped to WARN unless every namespace-year is simultaneously zero. ARTIFACTS.md is updated in the same commit as the emitting code and its validate assertion, per the repo's additive-contract rule. Tests cover the new checks plus an end-to-end publish.build() proof that headline fields stay blind to deficiență episodes.