ruff_python_spo: the config-consuming extractor and the promotion gate

[AdaWorldAPI]

Summary

Increment 3 of the self-adaptive drill loop, closing the boundary drill.rs's module doc has named twice now: "this does NOT build a promotion gate." One rule wired end-to-end — not all four residual reasons at once. A real config surface plus a real gate is worth more than a wide, partially-wired one.

PlainDrillConfig::unwrap_optional_annotation resolves T | None / None | T to T's own field_type reading. Chosen because it was the only cross-corpus candidate from #97's measurement (dismech 96, ruff/scripts 12, genuinely absent — not merely below threshold — on the A2UI SDK), unlike the LinkML call:PermissibleValue/call:EnumDefinition factories, which are corpus-scoped and would need a per-corpus row, not a generic one. field_type_from_annotation recurses into the non-None operand; a chained union (str | int | None) has a BinOp as its left operand, so the shape check fails and it stays unresolved rather than guessing which arm to keep. Off by default — PlainDrillConfig::default() is pinned byte-identical to the config-free path.

The gate

drill::ratify_optional_unwrap runs extraction with the rule OFF and ON over one source, then independently verifies the claim two ways that share no code with the resolver:

1. A separately-written shape test (is_shaped_t_or_none), applied by re-parsing the source and re-walking the raw AST directly — not by trusting field_type_from_annotation's own success/failure — counting how many of the newly-resolved sites are genuinely T | None shaped. Must equal the resolver's own count exactly, in either direction: the resolver claiming to fix something the independent check disputes is exactly as much a defect as missing one it should have caught.
2. Every resolved value is checked against a real-type-name shape (non-empty, alphanumeric) rather than trusted as "present, so fine."

Measured, real corpora, per-file (plain_ratify example)

dismech (84 files):      baseline=96 resolved=91 verified=91 ratified=YES
ruff/scripts (23 files): baseline=12 resolved=12 verified=12 ratified=YES
a2ui sdk (127 files):    baseline=0  resolved=0  verified=0  ratified=YES

dismech's 96 − 91 = 5 gap is the honest remainder: non-optional binop shapes (chained unions, int | str) the rule correctly declines rather than guesses on — the ratification proves the claim is exact, not that the rule resolves everything.

A falsifier finding worth recording rather than discarding

The first disable run on the independent verifier (is_shaped_t_or_none forced to always return true) reported the target test green under mutation — genuinely uninformative, not a false pass: the independent-verification loop only visits sites the resolver itself flipped, and a correct resolver structurally can never disagree with a correct independent check without a second bug existing first. Traced to the right level: RatificationReport::ratified()'s own disagreement-detection logic now has direct unit coverage (ratified_is_false_when_the_counts_disagree_in_either_direction, four hand-built cases), disable-run verified on that function directly (forced to true, the new test failed; restored, green). Two more disable runs on the resolver itself (optional_operand neutered; the config gate bypassed) both verified red-then-green normally.

A process note, for the record

Mid-build I reused the branch name from #97 (claude/plain-spo-config-drill) for this increment's git checkout -b; since that branch already existed (merged), the checkout failed silently and every edit landed on local main instead. Caught before push — origin/main was never touched — by noticing uvx prek run --from-ref main --to-ref HEAD reported "no files to check" on a branch with real changes. Moved the commit to this correctly-named branch, reset local main back to origin/main, and re-ran every gate on the corrected branch before this push.

Test plan

• cargo test -p ruff_python_spo — 84/84 (75 pre-existing + 9 new)
• cargo clippy -p ruff_python_spo --all-targets -- -D warnings — clean
• cargo fmt + RUSTDOCFLAGS="-D warnings" cargo doc --no-deps -p ruff_python_spo — clean (two intra-doc links de-linked, same pattern as prior PRs)
• uvx prek run --from-ref main --to-ref HEAD — all hooks pass, tree clean after
• plain_ratify output reproduced over all three corpora, pasted above

---

Generated by Claude Code

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_73a1f384-580a-482c-bafe-55924ecd46d2)

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.