Separate a session-created commit from the edits it contains

Traceability

• Spec ID: inspector-commit-provenance
• Status: Implemented
• Extends the evidence vocabulary of inspector-evidence-selection and the trace surfaces of inspector-trace-view; the correlation contract stays in commit-session-correlation.

Intent

Correlation could say a commit was near a session, but not that the session created it, and it could not tell a commit whose files were edited in the same session from a commit that only wrapped up pre-existing workspace changes. On a real local report that gap read as authorship: a session with zero observed Edit/Write paths still showed commits in its lane, and every commit of a session appeared related to every file-touching call in it.

The Inspector already retains the two facts needed to close the gap: an observed create-commit tool call with a start and duration, and Edit/Write calls with observed start times and exact repository paths. This slice reads the commit time against those windows so a reviewer can see which claim is being made.

Acceptance Scenarios

• AC-1: When a commit time falls inside an observed Create commit call window, correlation reports high confidence with the call id, and that match ranks above session-window candidates of the same confidence.
• AC-2: A commit created inside such a call carries the observed-commit evidence kind at observed strength, and its limitation states that its files may have changed before the observed session began.
• AC-3: An Edit/Write call links to a commit only when its observed start sits after the previous directly linked commit, at or before this commit, and one of its exact repository paths occurs in the commit, so a single edit cannot leak forward into every later commit in the session.
• AC-4: When a directly linked commit has no linked Edit/Write path, the workbench and the Evidence Drawer say so instead of implying the session authored the files.
• AC-5: Selection relates a commit to the call that created it and to the Edit/Write calls linked to it. Same-path overlap remains the fallback only for links that are not direct.
• AC-6: On a wall-clock axis, directly linked commit times render as focusable commit events that report their linked call and path counts, and selecting one selects the commit.

Non-goals

• Claiming authorship of committed lines from any observed call.
• Attributing a commit whose time falls in no observed call window.
• Changing the Feature Tree, Git, or native-session sources of record.
• Reconstructing edits the host never observed or privacy filtering removed.

Design Decisions

Provenance, not authorship. observedCommitCall matches the commit time against an observed create-commit call window with a 2 s tolerance on both ends. It proves only that the session created the Git object; the separate linkedEditEvidence pass answers whether the session was also observed editing those paths.

Bounded edit attribution. Edit linkage is windowed by the previous directly linked commit rather than by the whole session, and requires an exact changed path. Both bounds are needed: time alone repeats one edit across later commits, and path alone ignores commit order.

Direct links win the lane. Where a session has direct links, Date view, commit lanes, and related-object grouping use them instead of mixing in contextual same-path neighbours, which is what made every commit look related to every call.

Absence is stated. A direct link with no linked edit renders an amber commit bridge and an explicit fact, so the missing evidence is visible rather than inferred from an empty list.

Test and Review Evidence

• AC-1: a commit inside an observed Create commit call outranks broad session-window candidates asserts ordering, confidence, and the call id.
• AC-2/AC-4: report model distinguishes a session-created commit from files edited in that session asserts the evidence kind, strength, empty linked edits, and both the fact and limitation copy.
• AC-3: report model links Edit/Write calls to the next direct commit by time and exact path binds two commits in one session to one call each.
• Regression: npm test passes with these changes in place, including test/reporting/harness-inspector.test.mjs and test/sessions/commit-session-link.test.mjs.

Risks

• Correlation: the 2 s tolerance is a clock-skew allowance around an observed window, not a proximity heuristic; a commit outside every observed call keeps its previous candidate or contextual grade.
• Reading: an amber commit bridge marks missing evidence, not a defect, and says so in its own copy.
• Privacy: no new text or path is retained; commit events reuse stamps and paths the report already exposes.