Select native pytest keyword and marker scopes throughout generation and repair #46

Scope — estate-a3425ebf9874 / github_estate

Add native pytest keyword/marker selection to TestPilot's generation run: run --pytest-k EXPR --pytest-m EXPR and keyword-only TestPilot(..., pytest_k=None, pytest_m=None). The same literal native expressions apply to the existing baseline, generated verification and every repair. Both filters intersect through pytest itself. Selected authored passes alone never qualify a generated patch; generated-case admission remains unchanged.

The optional pytest_selection result context records literal keyword/marker values and appears in JSON, Markdown and HTML with an explicit statement that unselected tests were not verified. Planner/editor/repair prompts receive that context. Default behavior and absent-field compatibility remain unchanged. Saved-test recheck keeps its current full configured-suite behavior; it does not silently inherit run filters.

Actual original evidence

Original main 191cca4e416286a9e1fa4b5d5daf5fc30b936db6, complete tree 2602440d626872f0293755723dc6ca610b1e360b, was captured with all 91 relevant source/test/eval/config Git blobs verified. Native ThinkPad Python 3.14.4, pytest 9.1.1 and coverage 7.16.2 executed six original controls: ordinary generation succeeds, deliberate unselected authored failure causes default failure, both proposed CLI flags refuse with exit 2 and no report, while the existing low-level pytest runner successfully uses each native filter and admits an actual generated pass. Authored input bytes remain unchanged.

The full unchanged baseline records 389 passed, 18 subtests passed and 1 failure: an existing one-second deadline test did not reach its helper-start marker. Its one isolated unchanged control subsequently passed in 3.06 seconds. Both original outcomes remain retained; neither the timeout source nor test is in this scope. Witness receipt SHA256 7bd23268f54dd287ccca324e472f47931685b2251a225a01b56196dfc2398c69; full original log 6fd70efcfbc73ddcd7692b7fed19ce11d943f135cefcbcdc12fd8f837b2d95bd; isolated receipt 4f325ec70c839adcac3a6e2835f77f9fbb83f6ae425a340b36731201ed0d28a5.

Fresh main cec50d8df4499ba509615e179d82ef3861379bd2 adds PR42's raw Git-diff preference isolation. Its complete incoming delta was read and will be preserved, along with PR40's no-coverage behavior.

Narrow source fence and receiving

• Additive two run options and constructor forwarding in testpilot/__main__.py.
• Selection configuration/pass-through, result/context and Markdown reporting in testpilot/loop.py.
• Visible conditional selection context in testpilot/html_report.py.
• New focused real-pytest tests and docs/PYTEST_SELECTION.md, additive README usage, unique receiving evidence.
• Preserve sandbox/process custody, generated collector, diff/target selection, recheck, model/provider transport, dependencies, workflows and existing saved-report readers.

Active #35/#43 cancellation, #39 finder, #26/#33 comparison, #44 two-revision regression and #45 diff admission retain their scopes. Shared CLI changes are additive and will compose with current owners. No competing timeout repair will be attempted.

Native owned receiving root: /srv/hamon-estate/receivers/estate-a3425ebf9874-testpilot-pytest-selection; bounded native capture 9d1fcc586481e23416ae6e276c21bb99edafad07 is explicitly not upstream ancestry. Root has independently sealed 13 black-box consumer groups before candidate implementation and separately reproduced the existing default pass/missing-keyword refusal. Actual candidate execution, independent receiving, preserved true-parent source publication and normal hosted CI precede integration.