fix(indexer): skip unreadable files instead of aborting the whole index

[dkhokhlov]

A file that is unreadable at the OS level (e.g., an LSM denial returning EPERM
despite readable mode bits, or a permissions EACCES) aborted the entire indexing
fixes #296

Fix

Wrap both hashFile calls in try/catch:

• Main loop: record the file in stats.skippedFiles with a new "unreadable"
  reason and continue, so the remaining files index.
• Freshness check: return reason: "unreadable" (reuses the existing
  IndexFreshnessReason) so the next index_codebase rebuilds — which skips the
  unreadable file — instead of aborting.

Add "unreadable" to the SkippedFile.reason union in src/utils/files.ts.

Repro

~/SOURCE/Matlab2/lib/@medusa7/calc_spreads.m is LSM-protected (open() returns
EPERM errno 1, lsattr EPERMs, despite 0644 owner-owned perms — expected by
design). Its readdir position in lib/@medusa7 (219 .m files) is 110
(0-indexed), so it was skipped under the default maxFilesPerDirectory: 100 but
included once the global config raised it to 500, aborting ci -p with
Operation not permitted (os error 1) and 0 files processed. With this fix the
run completes and the file is reported as skipped; 973 of 974 .m files index.

Test

tests/skip-unreadable-file.test.ts indexes a project containing a readable file
and a chmod 000 file (EACCES — same catch path as EPERM); asserts the run
succeeds, the readable file is indexed, and the unreadable file appears in
skippedFiles with reason: "unreadable". Existing host-paths /
host-mode-paths suites stay green (41/41).

[Helweg]

Thanks for the focused fix. The two guarded hash paths look right. Could you add one regression test that exercises the freshness-check path as well? The issue affects both the main indexing loop and getIndexFreshness, while the current test covers only the indexing loop. Once that is covered, this should be ready for CI approval and merge review.

[dkhokhlov]

Added a regression test covering the freshness-check path.

The test builds an index first so getIndexFreshness passes the "missing" early return and reaches the hash loop, then calls getIndexFreshness() with the unreadable source file still present. It asserts { readable: false, current: false, reason: "unreadable" } rather than rejecting with the hashFile permission error.

Verified the test fails when the freshness guard is reverted (the pre-fix Permission denied throw aborts the check) and passes with it restored. Both tests in the suite are green.