fix(indexer): fingerprint chunkers at import time, pass language overrides as memo arg

[georgeh0]

Follow-up to #287 (issue #285). Fixes two bugs in how #287 detects chunking-config changes and makes the language-override path use its input directly.

Bugs fixed

1. The fingerprint could drift from the running code. fix(indexer): reprocess unchanged files when chunking config changes #287 hashed each chunker's source file from disk on every index run, but the daemon runs the code it imported at startup. If you edited a chunker and ran ccc index before a daemon restart, the old code ran and its chunks were memoized under the new fingerprint. After the restart the fingerprint was unchanged, so the stale chunks stayed for good.
2. Some chunkers re-indexed everything on every daemon restart. For a functools.partial or a callable instance, inspect.getsourcefile raises TypeError, and the repr(fn) fallback contains a memory address.

Changes

• Chunkers are fingerprinted when they are imported. Right after _resolve_chunker_registry imports a chunker, it wraps it in LoadedChunker(fn, spec, module_sha256): the "module:attr" string from settings.yml plus a sha256 of the imported module's file. The digest is cached per loaded module object. Without that, ccc reset or loading a second project in the same daemon would hash the edited file while importlib returns the module it already loaded, which reproduces bug 1.
• CHUNKER_REGISTRY is now change-detected (detect_change=True). Its value type changes from dict[str, ChunkerFn] to dict[str, LoadedChunker], and LoadedChunker is exported from cocoindex_code.chunking. LoadedChunker.__coco_memo_key__ returns (spec, module_sha256) and never the function. cocoindex fingerprints a function by module and qualname only, so a body edit wouldn't register, and lambdas and closures can't be fingerprinted at all. process_file already reads the registry, so no separate dependency is needed. Project.create(chunker_registry=...) takes the same mapping.
• Language overrides are a real process_file argument. indexer_main builds the suffix → language map once per run, and process_file uses it, so it is part of the memo key. The per-file load_project_settings call is gone. Override changes still take effect on the next run without a restart.
• Deleted chunking_fingerprint, _chunker_source_digest, the unused chunking_config argument, and the __all__ entry.
• README: the chunking-config paragraph now describes the actual behavior.

Behavior notes

• Granularity: any change to a chunker or to language_overrides re-chunks every file, not only the files that entry applies to. Chunks whose text didn't change reuse their cached embeddings, because embed is memoized. I checked this: after a fingerprint change the file is reprocessed but embed isn't called again.
• Limit: only the module file named in module: is hashed. An edit to a helper module it imports isn't detected (documented in the README).
• Upgrade: process_file's arguments changed, so the first ccc index after upgrading re-chunks every file once. Embeddings come from the cache.

Tests

All in tests/test_reindex_on_chunking_config_change.py. The existing two end-to-end tests still pass. test_chunker_change_reprocesses_unchanged_files now builds its chunkers with the daemon's _resolve_chunker_registry instead of passing bare functions.

New test	On main before this PR
Same config: second index run and daemon restart report all files unchanged (plain function)	passes (no bug for plain functions)
… same, functools.partial	fails: all 3 files reprocessed after restart
… same, callable instance	fails: all 3 files reprocessed after restart
Edit the chunker module, index before restart, restart, index → chunks reflect the new code	fails: stale v1 chunks after restart
… same, but the project is reloaded in the same daemon before the restart (ccc reset path)	fails: stale v1 chunks after restart. Also fails with the fix if the per-module digest cache is removed
Changing module: to a different callable in the same unedited module reprocesses the file	passes (#287 already covered spec changes; this guards the new fingerprint)

The duplicated _StubEmbedder from test_chunker_registry.py and this file is now a shared stub_embedder fixture in conftest.py.

Checks: uv run pytest tests/ (325 passed, Docker e2e deselected), uv run ruff check ., uv run ruff format --check ., and uv run mypy .. mypy reports only the 2 errors in scripts/find_best_models.py that already exist on main.

Also in this PR (CI fixes)

• Lint left on main by fix(daemon): guard socket unlinking on shutdown by checking bound node identity #291: ruff format and end-of-file fixes in daemon.py and tests/test_daemon.py. Every Pre-commit job on main failed at lint, which kept pytest from running.
• fix(daemon): guard socket unlinking on shutdown by checking bound node identity #291's socket inode test removed: with lint passing, pytest ran it for the first time, and it failed on ubuntu. It never ran daemon code, and on ext4 the replacement file reuses the unlinked inode number. fix(daemon): keep a displaced daemon from deleting its replacement's socket #300 removes it as well.
• Windows: the reindex tests now write exact-content files with write_bytes. On Windows, write_text produced \r\n, which failed fix(indexer): reprocess unchanged files when chunking config changes #287's test on main.
• Known, not fixed: on ubuntu 3.14t (free-threaded, continue-on-error), the in-process restart tests can hit environment already open in this program. Garbage collection doesn't free cocoindex's LMDB environment before the test reopens it, and cocoindex has no explicit close. fix(indexer): reprocess unchanged files when chunking config changes #287's restart test already fails this way on main.

Follow-up (out of scope)

INDEXING_EMBED_PARAMS in shared.py also lacks detect_change=True, so changing the indexing embed params (for example prompt_name) doesn't re-embed files whose content is unchanged.

🤖 Generated with Claude Code