Cover the localize text pipeline; raise the coverage gate to 88

[thequantumfalcon]

What

sts2/localize.py was 12% covered in CI and accounted for 39% of all uncovered code in the repo. The only test that exercised it ran against a real game install and skipped everywhere else, including CI.

That is the worst place in the app to have a hole. English descriptions never pass through this pipeline, so a game patch that reshapes a template silently ships half-rendered markup or a wrong number to every non-English player while the suite stays green.

It needs no game to test: of 970 lines only 16 touch the filesystem, and the parser, the English alignment pass and the renderer are pure text → text. tests/test_localize_pipeline.py drives all of them from fixture strings, including the full overlay build.

Template shapes in the fixtures were taken from the shipped English card, relic and potion descriptions of build 0.107.1, so they exercise forms that actually occur (NUM 881, PLURAL 173, ENERGY 106, SHOW 65, STAR 27, ENCH 3, COND 1) rather than invented ones.

Coverage: 74% → 90.48% in CI

Module	Before	After
localize.py	12%	94%
behavior.py	52%	100%
graveyard.py	61%	100%
config.py	69%	98%
__main__.py	71%	97%
fetcher.py	74%	88%
sources.py	77%	94%
app.py	81%	92%
updater.py	82%	91%
analytics.py	88%	96%
knowledge.py	86%	89%
ghost / rivalry / persist	71 / 79 / 81%	100%
prophecy.py	80%	97%

Most of the above was previously exercised only by mocks asserting mock.called, so each module could have returned a constant and stayed green.

--cov-fail-under moves 72 → 88, raised in three steps and each time from a measured CI run rather than an estimate.

Three bugs fixed, all found while covering the code

1. The overlay invented upgraded card descriptions. _build_all gave every no-alignment card an up variant whether or not English had upgraded text. card_detail.html renders an "Upgraded:" block whenever the field is present, so six cards — Apotheosis, Despair, Lantern Key, Sharp Edge, Wish, Reserves — showed translated players a section English readers never see, restating the base description back at them. Thirteen languages, so 78 entries.

Verified against a real German build off an installed game: all six clean, the 545 legitimate upgraded translations untouched, and nothing in the 608-card overlay claims an upgrade English does not have.

2. A duplicate "Upgraded:" block. Separately, 98 English and 120 German cards rendered an "Upgraded:" section identical to the description above it.

The obvious fix — dropping the field from the overlay — was measured and is wrong: knowledge.py skips absent overlay values, so the card falls back to the English upgraded string under a translated description. CARD.ANOINTED came out as a German description with 'Retain. Put every Rare card from your Draw Pile into your Hand. Exhaust.' beneath it, across 148 fields.

Fixed in the template instead, with the guard deck.js already applies to its popover. Measured: 490 English and 425 German cards keep their block, and no card whose upgraded text genuinely differs is hidden.

3. log.debug(" ", mm) passed an argument to a format string with no placeholder, which raises inside logging whenever DEBUG is on and the token-name check has anything to report.

Also removed an unreachable guard in detect_tilt (len(runs) >= 3 is checked above, and _group_sessions always returns at least one non-empty session for non-empty input).

The measurement gap has closed

This repo carried a ~7-point laptop-vs-CI coverage gap. It was almost entirely localize.py reading the installed game — 87% where the game exists, 12% in CI. Now that it runs on fixtures, both environments agree to within a point. The only remaining local-only difference is risk.py + diagnosis.py, which are gitignored and absent from a CI checkout. The stale comment in ci.yml documenting the old gap has been rewritten.

Two statistics validated against their definitions

Rather than against a recorded output, which would only pin whatever the code happened to do:

• _consistency_index against the closed form of the R/S statistic for a linear ramp: cumulative deviation is S_k = k(k−n)/2, so R = n²/8 exactly and S = √((n²−1)/12). Checked at n = 10, 20, 40, 100. (An initial guess that it approaches 1 was wrong — it is 0.773 at n=40 — and deriving it exactly is what caught that.)
• plural_index_ru against real Russian agreement, including the 11–14 exception: 1→0, 21→0, 101→0, 2/3/4→1, 22→1, 11/12/13/14→2, 111→2.

Verification

• pytest -q — 1570 passed
• pytest -q -m browser — 15 passed (required check, invisible to a plain run)
• mypy sts2 — clean
• ruff check sts2/ tests/ — clean
• All 15 CI checks green on Linux (3.11/3.12/3.13), Windows and macOS

Known, not addressed

The root cause behind bug 2 is upstream: the game's description templates omit the keyword prefixes (Retain., Innate., Exhaust.) that the wiki-derived English carries, because those live in card metadata. That affects base descriptions too, not only upgraded ones, and reconstructing them across 13 languages is a feature rather than a fix.

routes.py (87%) is the largest remaining coverage target.