Python abridging goldens #

These fixtures are unified diffs for the Python files changed by three real upstream commits that exposed different abridging failures. They are not claimed to include unrelated non-Python files from those commits. Each .plan.json is the model-authored source-coordinate portion of the plan; the compiler merges its deterministic mandatory import-removal plan before rendering. Golden plans must not target compiler-owned import rows themselves. Each .golden.diff is the exact machine-rendered reading diff.

• Django 526b1b414d8e215bf627b5722df12a09346dbf6b (June 8, 2026), “Refs CVE-2026-48587 -- Added helper to properly split header values.” Protects the token-list contract and quoted-value caveat, stripping transformation, representative call-site behavior, distinctive whitespace/wildcard/separator stimuli, required setup, and outcomes.
• Flask c17f379390731543eea33a570a47bd4ef76a54fa (February 18, 2026), “request context tracks session access.” Protects the property contract, backing-session lifecycle, save-path bypass, public type narrowing, proxy effect, and accessed/modified/Vary outcomes while collapsing repeated backing-field churn and assertion batches.
• pytest b4e846616cbb0ba74dc548f7066b09d820f5dc05 (July 22, 2026), “Apply warning filters to pytest_configure (#14760).” Protects symmetric evidence for the exact warning-filter move, plugin/lifecycle conditions, filterwarnings and hook-order stimulus, required pytester setup, and pass/no-warning outcomes.

The deterministic quality gates are absolute, so a snapshot cannot be made more verbose merely by claiming that it still compresses something:

fixture	raw changed	max changed	raw physical rows	max physical rows	raw bytes	max bytes
Django	178	89	329	155	14,468	6,685
Flask	124	63	234	120	8,803	5,037
pytest	109	42	159	65	6,468	2,615
Corpus	411	194	722	340	29,739	14,337

When revising a plan:

• retain contracts, security/compatibility caveats, behavior-changing conditions, transformations, effects, distinctive stimuli, outcomes, and any setup required to understand or execute the retained test scenario;
• after one representative anchor, remove purely mechanical rename/call-site hunks; remove default context, changelog prose, repeated fixtures/cases, and duplicate assertion spellings that add no new outcome;
• fold repetitive interiors only when the fixed ... preserves useful suite shape; never hide the owner, table/configuration referenced by surviving code, delimiter boundary, or lifecycle edge;
• leave import hiding to the compiler and preserve identical treatment for both sides of every detected exact move.

Regenerate snapshots only after reviewing the semantic assertions and budgets:

UPDATE_GOLDEN=1 go test ./meat -run TestPythonGoldenCommits

Do not raise a budget just to bless a larger snapshot; document the semantic anchor that requires the extra rows first. These compiler goldens should only be regenerated as part of deliberate corpus tuning, not routine rubric edits.

TestPythonGoldenCommits is the hard gate: it is hermetic and locks exact snapshot bytes, semantic anchors, move behavior, and the absolute budgets above. The opt-in TestRubric_PythonGoldenCommits is instead a costed, stochastic smoke test of the configured model. It requires unconditional import absence, a small set of stable semantic minima, and broad retention ceilings calibrated from recorded and review runs; it deliberately does not demand the hand-authored plan's exact folds, removals, or retention.