feat: per-adapter sensitivity evidence line (issue #68)

[geuben]

Summary

• Fixes the sensitivity check's observed: evidence capturing the wrong output line on some runners (xdist worker headers, XCTest console noise)
• Adds Verdict.target_evidence to each adapter: pytest extracts the first E-line of longrepr, xctest the first : error: line of the per-test window, gradle the first line of the <failure> message, vitest the first line of failureMessages[0], exec the last non-empty combined-output line
• Schema v7: sensitivity_check.evidence_line TEXT stored at check time; render prefers it over the raw first line; empty evidence → <no assertion line captured> sentinel; NULL (legacy rows) → old first-line fallback; lines over 160 chars are tail-capped with …

Commits

86db050 docs: friction log and PRD update for issue-68-sensitivity-evidence
146e153 feat: tail-keeping cap on the observed evidence line
e6d5235 test: an over-long observed line keeps its tail
a7004cc refactor: legacy rows with NULL evidence keep the first-line fallback
8e03088 feat: render the no-assertion-line sentinel instead of wire noise
8e68e09 test: empty evidence renders <no assertion line captured>
ddc6024 feat: friction log prefers evidence_line for the observed snippet
b2dc16f test: observed line shows evidence_line, not the raw first line
739ed8a refactor: sensitivity check stores the adapter's evidence line in the ledger
619c0bd feat: schema v7 — sensitivity_check.evidence_line stored at check time
53ad973 test: sensitivity check persists evidence_line on its ledger row
9012b32 refactor: exec evidence falls back to the last non-empty output line
78771ff feat: exec evidence falls back to the last non-empty line
1df55d5 test: exec evidence is the last non-empty combined-output line
bebee75 feat: gradle evidence extracted from the failure element's message
e0a5301 test: gradle evidence is the junit failure message line
248e550 feat: vitest evidence extracted from failureMessages[0]
3c0c019 test: vitest evidence is the first failureMessage line
1f1b40b refactor: xctest evidence is the error line, not interleaved console noise
5817c6a feat: xctest evidence is the first ': error:' line of the test's window
e04e009 test: xctest evidence line ignores console noise in the test window
dad64c6 test: pin empty pytest evidence when no E-line exists
b464cf7 refactor: pytest evidence is the assertion line, not the xdist worker header
a2b7ecb feat: Verdict.target_evidence — pytest extracts the first E-line of longrepr
a4be0c7 test: pytest evidence line skips the xdist worker header

Test coverage note

All decision-point operators in new code are covered by at least one test (pseudo-mutation analysis passed). 429 tests passed. Gates: uv run ruff check src tests (clean), uv run pytest (429 passed).

Plan fidelity

11/11 cycles declared and delivered, 0 skipped. 1 red_first_violation integrity event. 0 blockers, 0 human interventions.

Closes #68