v4.5.3

Added

• PARTIAL_BASELINE trust state (#276). When the diff JSON reports
  partialBaseline / explainedMissingBaselineSurfaces, the Action classifies
  ledger-explained missing baseline surfaces as repair debt — after residue,
  inventory, and certification failures, before DEGRADED_BASELINE and visual
  approval. changed stays false for baseline-only gaps; mixed greenfield runs
  still exit 3 but PARTIAL_BASELINE outranks VISUAL_APPROVAL_REQUIRED
  while explained gaps remain. Commit status and job block fail with repair-base
  guidance; the approval box cannot clear this state.

• styleproof-ci ephemeral worktrees (#277). Restore probes and cold base
  install/capture run in detached git worktrees under RUNNER_TEMP (or the OS temp
  dir), so the consumer checkout is never checked out to --base. Head capture may
  still run in the consumer at --head. Worktrees are always torn down on success
  and failure; invalid SHAs fail loudly before capture.

• Canonical comparison truth (assessComparisonTruth). Diff JSON, durable
  report, PR comment, and trust state share one assessment of raw vs reviewable
  findings so they cannot invent disagreeing verdicts.

Fixed

• Baseline @auto crawl failures match head capture widths in diff/report (#276).
  Viewport-detection failures recorded as surface@auto now classify head-only
  surfaces as broken baseline capture (not greenfield new) for any width of that
  exact surface key; width-specific ledger entries stay exact (about@1280 does
  not match about@900). Diff exit 3 applies only to genuinely new surfaces.
  Report Markdown escapes inline injection in failure reasons; manifest JSON keeps
  the original failure ledger keys.

• Partial baseline capture on the base branch (#276). styleproof-map --tolerate-surface-failures (and styleproof-ci on cold base capture only)
  records per-surface failures in the manifest, publishes a partial bundle when at
  least one map succeeds, and leaves self-check / zero-map / head capture
  fail-closed. Diff and report surface repair-base warnings instead of treating
  ledger-explained surfaces as greenfield new.

• Report headline and global-chrome banners now count surface bases consistently.
  The summary line reports unique changed product surface bases and, when more
  capture keys exist than bases, adds a labeled variant count. Product bases prefer
  authoritative capture metadata.surfaceKey (live-state and popup expansions share
  their declared base); older captures without metadata still strip only @width.
  The global-chrome tier names captured surface bases (not raw variant keys). A
  one-line glossary beside the headline explains @width and state/popup keys as
  variants of a base.

• Diff/report/verdict coherence for derived-only (reflow) deltas.
  styleproof-diff can report raw computed-style differences while
  generateStyleMapReport strips size/position longhands and previously claimed
  "all surfaces identical" with no crops. That pair must never become
  VISUAL_APPROVAL_REQUIRED (no reviewable evidence). Both CLIs now emit
  reviewableCounts + reportConsistency; raw-only noise fails closed as
  CERTIFICATION_FAILED (action gate + status), the report headline names the
  consistency failure instead of "identical", and approval UI stays off.