refactor(git): migrate guarded workflows to Python

[BjRo]

Why

Closes #181. Darrow Git’s substantial Bash mechanics need native Linux, macOS, and Windows support without weakening mutation safeguards or exact publication evidence. This PR also fixes #171 and #174 in separate commits before the migration.

What changed

• c2c4258 bounds observation of delayed PR-head propagation without repeating publication; 326f6bb keeps successful user-facing reports concise while preserving the complete machine record.
• Move all five workflows to one contained Python + UV package with frozen resolution, argument-vector Git/GitHub calls, and native filesystem handling. Remove all six legacy Bash entrypoints; callers use the frozen UV commands directly on every platform.
• Preserve worktree, index, hook recovery, duplicate prevention, exact-head verification, and evidence reconciliation contracts; add real-Git integration/property tests and fresh-copy platform jobs.
• Register the package, update invariants and usage, bump both Git manifests to 0.8.1, and teach eval fixtures to mount the backend, manifest-local resources, and writable UV scratch. Retarget the seven shell regression suites and dependent composition fixtures to the Python commands; the ticket-to-PR fixture patch bumps its manifests to 0.4.1.
• Simplify internal argument consumption, move commit dispatch into the CLI to remove the circular remediation dependency, and share identical provider-error and local-default-branch helpers. Preserve the outward interface and publication state model.
• Route Claude Git requests through their matching skills before preflight using a static native SessionStart reminder. This fixes skipped activation for clean trees, missing inputs, conflicts, and hook failures, including a reproduced history rewrite when the commit skill was skipped. The hook performs no Git/GitHub operations, prompt classification, or orchestration.

Verification

• bun run check:python: passed; Git package has 158 passing tests, 98.54% statement coverage and 95.86% branch coverage.
• Parser simplification matched all 31,866 pre-refactor outcomes, including exact refusal text/status and unchanged input lists; durable CLI tests check stdout/stderr routing and diagnostic precedence.
• bun run lint, lint:ts, lint:shell, typecheck, check:decisions, and check:docs: passed.
• bun test evals/runner/environment.test.ts evals/runner/fixture.test.ts: 14 passed.
• All seven retained shell regression suites passed locally against direct UV commands with Bash 3.2, including after internal simplification. Final platform CI passed on 8882332: Linux/macOS/Windows fresh installation, Python 3.10–3.13 quality, and Bash 5/3.2 regression runners (73 Python/platform jobs plus documentation).
• The two affected composition-fixture suites have 16 passing tests, including rejection of stale publication evidence and execution of guarded hook remediation through the copied Python package.
• Full migration/routing evaluation covered all 48 cases on each host; Claude had 49 successful trials, including one extra ambiguous-hook case selected by the substring filter. After wrapper removal, the conventional-commit case passed again on both hosts; after internal simplification, the hook-retry-authorization case passed on both hosts. Trials use Codex gpt-5.6-terra/medium and Claude claude-sonnet-5/medium, one trial per case per invocation, threshold 80%. These samples are not a reliability estimate.
• Matched pre-migration controls reproduced Claude's skipped activation for clean trees and its raw-Git amendment of history. With the context hook, both cases select the commit skill and satisfy the observable repository-state checks. Missing-token, conflict, ambiguous-hook, and all five nonactivation cases also pass.
• Fresh-context migration, Claude routing, wrapper-removal, and internal-simplification reviews completed; material migration findings were repaired, and the latter three reviews found no material findings.

Review notes

The runtime interface intentionally breaks compatibility with legacy shell paths; use the documented frozen UV entrypoints. Review the propagation and reporting fixes separately from the migration, Claude routing, and wrapper removal. Provider subprocesses use argument vectors; the existing explicitly authorized hook-diagnostic command remains the documented exception, executed by POSIX sh or native Windows cmd with the saved HEAD/index guard. Focus review on snapshot restoration and evidence identity/reconciliation. The Claude hook uses an executable plus argument array and is absent from Codex's manifest; fresh-copy CI exercises that command on each platform.

Checklist

• I have read and followed CONTRIBUTING.md, including the contribution
  licensing terms.
• I added or updated the applicable invariant before implementation, or
  this change does not affect a capability invariant.
• I added or updated colocated evals, or this change does not affect skill
  behavior.
• I confirmed that each changed plugin remains self-contained, or this
  change does not affect plugin content.
• I ran bun run check:python, or this change does not affect registered
  Python packages or their repository quality infrastructure.