perf(update): skip redundant basis re-hash across sparse chain steps

[peters]

Purpose

Autoresearch round 8 (update surface). Each chain step verifies the
patched file's target SHA-256; the next step's basis check then
re-reads and re-hashes the very same content. For the canonical sdk
step (one 4 KiB page in a 1.06 GB file) that is a ~0.55 s full-file
read + SHA-256 per step.

The chain walker now carries a VerifiedFileHashes map (path →
verified sha256) through apply_target_deltas:

• PatchFile: basis verification skipped when the cache records the
  expected hash; first step and any mismatch still verify fully
• WriteFile: target hash recorded after verification
• Delete: entry dropped

Behavior impact

• No change to patch formats, wire bytes, or the installed payload.
• Verification semantics preserved: every file state is still
  hash-verified before it is used as a basis (just not twice); the
  post-patch target hash and the per-step full-archive SHA-256 are
  unchanged. A tamper-between-steps regression test corrupts the file
  after step 1 (bypassing the cache) and asserts step 2 still fails
  with a hash mismatch — verified the skip path is actually taken
  (marker in the skip branch fired in the test).
• Single-shot apply passes a fresh empty cache (no cross-step state).
• Doc note: also restores the identity-chunk (format v2) block in
  docs/performance/update-chains.md, which was lost when a
  multi-block edit was rejected wholesale in the perf(diff): identity-chunk markers in chunked patches (format v2) #256 round, and
  records the corrected per-step breakdown.

Test evidence

• New regression test
  test_in_place_chain_with_verified_cache_detects_external_modification;
  existing carried-chain equivalence test updated to use a shared
  cache (asserts byte-identical output with caching on).

• cargo test --workspace green; clippy all-targets + strict
  (-D clippy::unwrap_used -D clippy::expect_used) clean; fmt clean;
  maintainability guardrail clean.

• Same-session A/B, 100-delta chain, scale 1.0, sdk-only, seed 42
  (baseline = main bba9f4f in a separate worktree, interleaved run):

  apply	baseline	with carry
  run	216,777 ms	156,599 / 158,057 ms

  −27.4% (0.9% spread); per step ~2.2 s → ~1.6 s; download bytes
  (503,524 B) and install tree (1,214,024,029 B) identical across
  all runs.

• Full-chain history on this benchmark: 657 s → 489 s (perf(update): carry the extracted tree across sparse chain steps #252) →
  221 s (perf(diff): identity-chunk markers in chunked patches (format v2) #256) → ~157 s (this change) — cumulative −76%, payload
  byte-identical at every stage.