fix(detect): detect Office source edits and stop re-parsing unchanged files (#1649, #1656)

[TPAteeq]

Fixes #1649
Fixes #1656

Summary

#1649 and #1656 both live in convert_office_file and pull in opposite
directions, but a single change — a source-content fingerprint — resolves both.

• Office files (.docx/.xlsx): modified sources never re-enter --update (sidecar keyed on path hash, source content never checked) #1649 (Office edits never re-enter --update): the sidecar name was keyed on a
  hash of the source path, never its content, and an early-return on
  out_path.exists() (the detect() creates duplicate converted .md files on macOS due to NFC/NFD Unicode path normalization inconsistency, causing --update to re-extract all Office files on every run #1226 no-churn fix) meant a modified .docx/.xlsx was
  never re-converted. The graph silently froze on the stale version, with no error.
• Perf: cache word_count in manifest — detect_incremental re-parses every PDF/docx on each run #1656 (re-parses every file every run): convert_office_file fully parsed the
  binary before that early-return, so Office files were re-parsed on every run even when
  unchanged; PDFs were likewise re-parsed by count_words → extract_pdf_text on every
  run, with no caching.

The single-fingerprint approach

convert_office_file now fingerprints the source by its raw bytes (md5).
Hashing the raw bytes does not unzip/parse the OOXML container, so it stays cheap —
that is the whole point. The fingerprint is recorded in the sidecar header:

<!-- converted from report.docx | source-md5: 3b1e… -->

On a later run:

• Sidecar exists and the stored fingerprint matches the source → return it
  without parsing (satisfies Perf: cache word_count in manifest — detect_incremental re-parses every PDF/docx on each run #1656) and without rewriting (so an unchanged
  source never churns its mtime, preserving the detect() creates duplicate converted .md files on macOS due to NFC/NFD Unicode path normalization inconsistency, causing --update to re-extract all Office files on every run #1226 guarantee).
• Sidecar missing, or the fingerprint differs (or is absent — a legacy sidecar) →
  re-parse and rewrite. The rewritten sidecar's new mtime/md5 then flows through
  detect_incremental automatically, so the edit re-enters --update (satisfies Office files (.docx/.xlsx): modified sources never re-enter --update (sidecar keyed on path hash, source content never checked) #1649).
  No other file needs to change for change-detection to pick it up.

This threads the needle that a naïve "always re-convert" would miss: it neither freezes
edits (#1649) nor re-parses unchanged sources (#1656/#1226).

PDF / word-count re-parse (#1656)

Word counts are now cached in the manifest entry and reused for unchanged files, so
PDFs are no longer re-parsed for their word count on every incremental run:

• save_manifest seeds/refreshes each entry's word_count, reusing the previous
  value whenever the content hash is unchanged and only (re)parsing a genuinely
  new/changed file (video files are skipped, mirroring detect()).
• detect() accepts the previous run's per-file counts and reuses the cached count for
  any file whose mtime is unchanged instead of re-parsing.
• detect_incremental loads the manifest first, hands those cached counts down to
  detect(), and proceeds as before.

total_words stays correct in every path (unchanged files contribute their cached
count; changed files are recomputed), so the benchmark command's corpus_words
keeps working — including via the skill's --update flow, which propagates
total_words into .graphify_detect.json.

Backward compatible: legacy manifests without word_count load unchanged (the existing
_normalise_entry dict passthrough preserves the new field, and old entries simply
recompute once on the next run).

Tests

Added to tests/test_detect.py:

• Edited vs. unchanged .docx and .xlsx via convert_office_file — asserts the
  converter re-parses after an edit and is not re-invoked when the source is unchanged.
• Legacy (pre-fingerprint) sidecar is upgraded once, then reused.
• End-to-end: an edited Office source re-enters detect_incremental().new_files, while
  an unchanged one stays in unchanged_files.
• An unchanged PDF is not re-parsed (extract_pdf_text not called again) on a second
  incremental run.

Test results

uv run pytest tests/test_detect.py tests/test_incremental.py tests/test_office_limits.py tests/test_manifest_ingest.py
# 153 passed

Full suite is green except one pre-existing, environment-only failure
(test_extract.py::test_collect_files_skips_hidden) caused solely by running inside a
.claude/worktrees/… checkout — it fails identically on the base commit with this
branch's changes stashed and is unrelated to this diff.