Follow-up to #63: measured provenance for worktree writes, branch -D / rebase / checkout --, floor calibration

Start by reading #63, commit 31d5ea7, and tests/provenance.test.ts to understand the landed force-push provenance and required regression coverage. Run the tests under tests/ with jev-latest and DEFAULT_JEV_POLICY; done means live before/after matrices for slices C–E, preserved force-push verdicts, cache-key and state-contract updates, a green suite, and clean tsc output.

Follow-up to #63: measured provenance for the remaining git classes (worktree writes, branch -D / rebase / checkout --, floor calibration)

Carries the slices of #63 that commit 31d5ea7 deliberately left out. The force-push half of #63 (proposals A-push + B) landed and #63 is closed, but its root causes 3, 4 (rebase/branch -D rows), and 5 are still unmeasured, and the criteria text for writes_outside_working_directory is still path-shaped.

Source: #63's own measured table (the rows that did NOT move)

Proposed slices, in the order #63 ranked them

C. Worktree-aware writes_outside_working_directory — the gate measures at classification time: is workingDirectory a linked worktree, what is the main checkout's path, and what paths did this session itself create? Writes inside the session's worktree, the parent repo's worktree directory, and gate-created paths are inside the workspace. Reading elsewhere stays a read. Risks: path-prefix matching needs it right (a main checkout contains the worktree dir), and the measured field must be gated to avoid re-introducing the injection surface #63's proposal A called out — measured by git plumbing, never author-supplied.

D. Branch-provenance measured fields for delete/rebase/checkout — same trust tier as gitPushProvenance: branch -D target merged into origin/main or HEAD?; cleanliness + stash presence for checkout -- .; ahead/behind vs the tracked upstream for rebase origin/main. Criteria read the fields by name (the lesson from the force-push landing: the model must be told where provenance lives, or the +-syntax FP shape recurs).

E. Floor calibration, only after D — #63 measured ordinary-but-consequential work converging on safe 0.65-0.73 / conf 0.47-0.53 against floors 0.8 / 0.5. If D's criteria give the model something to read, probabilities move for the right reason and floors stay put (#63 explicitly declined a global safeMinProbability drop). This slice is a measurement pass first, not an edit.

Acceptance (inherited from #63's discipline)

1. Each slice lands with a live measured matrix (before/after) on jev-latest + DEFAULT_JEV_POLICY, using real git plumbing in a controlled fixture, not synthetic state.
2. Zero regression on #63's force-push table: the three measured shapes from 31d5ea7 (no-op push, forward-only lease push, diverged push) must hold their post-fix verdicts.
3. New fields ride the session cache key and are labeled as gate-measured in the state contract.
4. Battery hash moves are expected and per-design (cache invalidation).
5. Tests under tests/ mirroring the matrix; suite green + tsc clean.

Provenance links

• #63 (closed) — root causes 1-5, the measured table, proposal A-E, and the open-question resolutions (no-op push = SAFE with no dialog; unpublished-branch lease push = UNSURE-able once measurable).
• 31d5ea7 — force-push slice: measureGitPushProvenance, gitPushProvenance, criteria text reading it by name; tests/provenance.test.ts.
• Open question inherited from #63: which battery questions cite each new field by name. Every new field gets its citing criterion enumerated in the PR, not discovered by review.