…ommitted changes

Four coordinated changes to fix the user-reported failure mode where
`entire review` would silently scope reviews against random
recently-merged feature branches (e.g., `origin/soph/add-codex-post-tool`)
instead of mainline, producing reviews dominated by 30+ unrelated
upstream commits.

1. scope.go: delete the "merged-into-HEAD ancestor with most recent
   committerdate" heuristic. It was designed for stacked-PR workflows
   but routinely picked unrelated branches because `git fetch` mirrors
   all of origin's branches by default and merged PR branches often
   live on for a while before deletion. Replaced with the existing
   fallback chain origin/HEAD → origin/main → origin/master → main →
   master. Removed slowDetectScopeBaseRef, isAncestorOf, and the
   candidate selection loop — ~130 lines of code go away.

2. cmd.go: register `--base <ref>` flag for explicit override (stacked
   PR review is now an explicit user choice, not an inferred heuristic).
   Validates the ref via `git rev-parse --verify <ref>^{commit}` before
   spawning agents; an unknown ref errors immediately with a message
   naming the bad ref. Threaded baseOverride through runReview →
   runSingleAgentPath / runMultiAgentPath → detectScope →
   ComputeScopeStats.

3. prompt.go: scope clause now reads "review the commits unique to
   this branch vs <ref>, plus any uncommitted changes in the working
   tree." Without this, agents (correctly) followed the prior
   commits-only wording and silently skipped uncommitted work — the
   most common case when a developer is mid-feature. Checkpoint
   context is preserved as a separate prompt section.

4. scope.go countFilesChanged: switch from two-dot diff
   (`git diff base..HEAD`) to three-dot (`git diff base...HEAD`).
   The two-dot form shows the tree diff between base and HEAD, so
   files modified on mainline AFTER the branch was cut appear as
   reversed deltas, inflating the banner's file count. Three-dot uses
   the merge-base, excluding upstream-only changes. Caught by codex
   during the smoke test review of this very PR.

Tests:
- TestDetectScopeBaseRef_PrefersMainOverAncestorBranches: replaces
  the obsolete TestDetectScopeBaseRef_ClosestAncestorPreferred which
  asserted the opposite (now-deleted) behavior.
- TestComputeScopeStats_BaseOverrideUsed /
  _BaseOverrideUnknownRefErrors /
  _EmptyOverrideUsesMainlineDetection: unit-level coverage for the
  override path.
- TestComposeReviewPrompt_ScopeIncludesUncommittedChanges: asserts
  the new scope wording.
- TestReviewCommandSmoke_BaseFlagThreadsThroughToPromptAndBanner /
  _BadBaseRefErrorsBeforeAgentSpawn: end-to-end command-level tests
  that catch wiring bugs (would have caught the silentErr
  suppression bug I had to fix mid-smoke).
- Deleted TestIsAncestorOf (function is gone).

Smoke-verified end-to-end on this branch: codex's review correctly
identified `Reviewed scope: no commits unique to origin/main; findings
are from the uncommitted worktree.` — proof that the right base was
picked, uncommitted changes are in scope, and the wrong-branch-review
failure mode is structurally fixed.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: 5600c1e3f60e

- cmd/entire/cli/agent/codex/reviewer_test.go: update scope-clause
  assertion to match the new wording. Was the only test outside
  cmd/entire/cli/review/ that asserted the old "review only the
  commits unique to..." string; missed it in the first sweep.
- cmd.go: --base help text and flag description now list the full
  fallback chain (origin/HEAD → origin/main → origin/master → main →
  master), aligned with the implementation in fallbackScopeRef.
- scope.go: ComputeScopeStats doc comment now mentions the `^{commit}`
  peel applied to the rev-parse --verify check, matching the actual
  code.
- scope_test.go: remove stale comment block that still described the
  old "closest ancestor preferred" behavior above the renamed test.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: 4e5cb1b17589

Cursor Bugbot caught an inconsistency in detectScope: when git.PlainOpen
fails AND baseOverride is set, the function silently returned ("", nil)
and the run proceeded in degraded mode — dropping the user's explicit
--base flag without warning. That contradicts the fail-loud contract
the function's own doc comment establishes for invalid overrides on
the ComputeScopeStats path a few lines below.

Now mirror the same pattern at the PlainOpen branch: on PlainOpen
failure with a non-empty baseOverride, return an error naming both
the override and the worktree path so the user can diagnose.
Auto-detection failures (no --base) still degrade gracefully.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: c21ae2d65511

Extends the existing checkpoint-context pipeline to surface
uncommitted agent work alongside committed checkpoint summaries.
Without this, reviews of branches with only uncommitted work
(common case: developer mid-feature) lose the rich context that
makes Entire's review better than vanilla agent reviews — agents
see the diff but no signal about what each session was for.

reviewCheckpointContext is now a thin composer over two sources:

  reviewCheckpointContext(ctx, worktreeRoot, scopeBaseRef)
    = committed-section + in-progress-section
    where:
      committed-section: existing "Checkpoint context from commits
        in scope:" block, sourced from entire/checkpoints/v1
      in-progress-section: NEW "In-progress session context
        (uncommitted):" block, sourced from session state files
        (.git/entire-sessions/<id>.json) + the on-filesystem
        prompt.txt that lifecycle.go appends to every turn

Inclusion criteria for in-progress sessions (reviewSessionContext):
  - state.WorktreePath == current worktree (canonicalised via
    filepath.EvalSymlinks so /var vs /private/var on macOS matches)
  - state.BaseCommit == current HEAD (work since last commit only)
  - !state.FullyCondensed (still in flight)
  - state.Kind != KindAgentReview (don't include the review agent)

Per-session entry format mirrors the existing prompt-fallback path
(reviewPromptText), so the agent-facing output is consistent
whether the work is committed or in-progress:

  <sessionID[:8]> <AgentType> [(touched: N file(s))] prompt: <latest>

Reuses reviewPromptText for prompt rendering — no new truncation
logic, no new caps. Sessions whose prompt.txt is missing or empty
are skipped silently.

The compose layer (joinReviewContextSections) emits non-empty
sections joined by a blank line; ComposeReviewPrompt already
skips empty sections cleanly, so no plumbing change needed for
the "0 commits + 0 sessions" degenerate case.

Smoke-verified end-to-end on this branch: built binary, ran review
with a fake codex that captures stdin, observed the section in the
captured prompt with live data from the current claude-code
session (correct session ID prefix, agent name, touched file count,
and latest prompt from the actual conversation).

Tests added:
- TestReviewSessionContext_IncludesActiveSessionWithLatestPrompt:
  unit-level, asserts the section text contains the session ID
  prefix, agent display name, and latest prompt.
- TestReviewSessionContext_SkipsSessionsOutsideScope: four-case
  exclusion check (FullyCondensed, WrongWorktree, WrongBaseCommit,
  KindAgentReview).
- TestReviewCommandSmoke_IncludesInProgressSessionContextInPrompt:
  end-to-end through the cobra command, asserts the in-progress
  block reaches the captured agent prompt. Catches wiring
  regressions between reviewSessionContext, the deps bridge, and
  ComposeReviewPrompt that the unit tests alone cannot.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: d08b2e17d004

Soph pointed out the doc comment still referenced `git rev-parse
--verify <ref>^{commit}` after the validation switched to
`repo.ResolveRevision` in b22552c. Update the comment to
match the actual code.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: e31eb25c73e6

… message, regression test)

- Sweep "closest ancestor" framing and dead git for-each-ref reference
  from 5 sites (prompt.go, scope.go, types/reviewer.go, CLAUDE.md) left
  behind by the heuristic deletion.
- Set cmd.SilenceUsage = true on the WorktreeRoot and currentHeadSHA
  prerequisite branches in both runSingleAgentPath and runMultiAgentPath
  so prerequisite failures no longer dump cobra's usage block at users.
- Improve the fallback-chain failure error in fallbackScopeRef: name the
  refs tried and point users at --base <ref> or `git fetch`.
- Add TestCountFilesChanged_ThreeDotIgnoresUpstreamOnlyChanges — a
  regression test where main advances past the branch point, so the
  three-dot vs two-dot diff distinction is observable. Existing tests
  use fast-forward branches and would pass under either dot count.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Entire-Checkpoint: 945292726919