feat(gre): desktop scoring adapter — synced gre_scorecard on dashboard open (Task 6)

[f15cubing]

What & why

Desktop-authoritative scoring adapter (Task 6). On GRE dashboard open, compute the three separate scores and write a synced gre_scorecard to col.conf so AnkiDroid can render it read-only (Task 7). Memory = FSRS-mastery Wilson range (None at n=0); Performance = not_available (no desktop attempt bank yet); Readiness = gated off with reasons + evidence panel — deliberately NOT derived from Memory (firewall / no-blend ceiling), never a bare number. Extends the W2 dashboard + the pure-stdlib outer-repo scoring/ package (imported via a sys.path bridge to the repo root).

Area(s) touched

• qt/aqt/gre/scoring_adapter.py (new) — compute_and_write_scorecard(col)
• qt/aqt/gre_dashboard.py — _write_scorecard(mw) on-open hook (guarded; never blocks the dialog)
• qt/tests/test_scoring_adapter.py (new) — 3 tests

engine/Rust PR? no — Qt-UI Python only; engine lane solely because it lands in the anki submodule. Docs ship in the companion f15cubing/speedrun PR.

Docs updated

Companion speedrun PR: docs/codebase/qt.md (§ scoring adapter), docs/codebase/INDEX.md, scoring/scoring.md; new Last-verified f15cubing/anki@ed74d1f.

Test evidence

Env: PYTHONPATH=pylib:out/pylib:out/qt ANKI_TEST_MODE=1 out/pyenv/bin/pytest -p no:cacheprovider qt/tests

• test_scoring_adapter.py → 3 passed (RED→GREEN: schema/persist/gated on a fresh collection · evidence panel present when gated · dashboard-open hook writes the card)
• All GRE aqt tests (test_gre_* + scoring_adapter) → 32 passed (no regression)
• ruff check + ruff format --check + mypy → clean
• Headless round-trip smoke: scorecard persists across close/reopen (no corruption); Readiness gated with reasons [<200 graded reviews, <50% topic coverage, no exam attempts yet]

Engine/Rust extra gate

N/A for the Rust-test items — no Rust/proto/FFI/scheduler change (Qt-UI Python only).

• Undo: preserved — col.set_config(..., undoable=False) creates no undo entry.
• No corruption: no schema change; round-trip open/close verified.
• Read-only invariant: the mastery read path is unchanged and still returns plain rows (never OpChanges); the adapter's only write is the gre_scorecard config value.
• Files touched: 3 (2 new + 1 hook line). Merge difficulty: low (isolated to qt/aqt/gre).

Made with Cursor

[f15cubing]

Engine-lane review — Task 6 desktop scoring adapter

Reviewed by the engine-lane reviewer (different agent than the builder). Recommendation: APPROVE — clear to merge once the merge-order/pin steps in note 2 are handled.

Note: submitted as a review comment rather than a formal --approve because GitHub's review API refuses approve/request-changes on a PR authored by the same account (Can not approve your own pull request). The approval recommendation stands. Not merging, per instructions.

What I actually reviewed: the real change is the tip commit ed74d1fe6 (parent 277f12e) = 3 files, +204:

• qt/aqt/gre/scoring_adapter.py (new, +118)
• qt/aqt/gre_dashboard.py (+18, guarded on-open hook)
• qt/tests/test_scoring_adapter.py (new, +68, 3 tests)

The gh pr diff spans ~60 files only because main is far behind this stacked branch (see note 2) — not scope creep in this task.

Base gate — PASS

• Scope — PASS. Tip commit is exactly the 3 expected files; no unrelated edits.
• Tests — PASS. Re-ran in the pre-built worktree: pytest qt/tests/test_scoring_adapter.py test_gre_dashboard_data.py test_gre_dashboard_mediasrv.py -> 17 passed (0.52s). Builder's posted green (3 adapter + 32 GRE aqt, ruff/format/mypy clean, headless round-trip smoke) trusted + spot-checked per building-and-testing.
• Docs — PASS. Updated in the companion f15cubing/speedrun#28 (qt.md § scoring adapter, INDEX.md, scoring.md; Last-verified -> ed74d1f).
• PR body — PASS. All template sections present.

Not a Rust change — CONFIRMED

Tip commit touches no rslib/, .proto, or pylib FFI — Qt-UI Python only; engine-lane solely because it lands in the anki submodule. Rust extra-gate items (>=3 Rust tests, etc.) correctly N/A.

Ceiling spot-check (read the code) — PASS

1. Undo preserved. Only collection write is col.set_config("gre_scorecard", card); set_config's default is undoable=False (pylib docstring: creates no undo entry, preserves undo history). No other mutation.
2. Read-only / no corruption. col.mastery_query read path is untouched by this commit and still returns plain rows, never OpChanges. Adapter only READs mastery + the W2 view-model and WRITEs the single config value; no schema change.
3. Three scores never blended; Readiness never a bare number. scorecard.build() nests memory/performance/readiness as separate objects. Readiness is shown=False, estimate=None with non-empty reasons (give_up() -> <200 graded reviews, <50% topic coverage, plus no exam attempts yet at n=0) and the full evidence panel (coverage_pct, confidence, reasons, best_next_topic). It is derived from review/coverage proxies, not from Memory — the adapter never feeds vm["memory"] into readiness and deliberately avoids rd.project(). Firewall intact.
4. No ETS leak. Adapter reads taxonomy + FSRS mastery rows + the W2 view-model only; never imports exam_items.json / eval/bank / any ETS anchor table.
5. AGPL headers present on both new files.

Non-blocking notes

1. (nit) The adapter relies on the default undoable=False instead of passing it explicitly as the PR body states. Correct today; passing it explicitly would match the body and be robust to any future default change.
2. (merge hygiene) PR is CONFLICTING and diffs ~60 files vs main because the branch is stacked on the not-yet-on-main W1 mastery / W2 dashboard / exam-mode / autoimport work. At merge, ensure that prior work is already on main so a squash of this PR contains only Task 6, then reconcile ankitects#28's pin to the squash-merged SHA (already called out in both PR bodies).