fix(autoindex): --no-graph durable path indexed zero documents for every code file - #316

Summary

On the durable-generation path (--no-graph), every source-code file was silently dropped: walked, sniffed correctly as family: code, counted, and then prepared as zero documents. The generation still committed and the run reported success (exit 3), so the loss was invisible — the #204 class of accepted-and-ignored input.

Root cause

prepare_artifact sniffs with the logical path (sniff_with_name(snapshot_blob, &source.path)) but extracts with only the blob path. code::extract was the one family extractor that recovered a parameter — the language, via path.extension() — from its content path instead of from Sniffed. Snapshot blobs are content-addressed and extensionless (sync-snapshots/<tx>/blobs/00000000), so the registry lookup never matched and every code file, in every language, became junk.

Fix

Sniffed now carries logical_name, set by sniff_with_name, and code::extract resolves both the grammar and the document title from it, falling back to the content path (which keeps the legacy path byte-identical: there logical == content). The title matters too — without it the fixed durable path would have indexed every code doc titled by blob ordinal 00000000.

Reproduction (run at HEAD, 803c85b)

The new regression test preparation_extracts_code_from_extensionless_snapshot_blobs fails on the parent commit exactly as the issue describes:

assertion `left == right` failed: a code file must prepare one document, not silent junk
  left: (0, 1)      # (records, junk)
 right: (1, 0)

End-to-end with this fix (release binary, throwaway server on :9483), the issue's own repro corpus (app.py + notes.txt) under --no-graph:

generation 1 committed — 1 datasets, 2 records live
{'ax_path': 'app.py',    'ax_format': 'py',  'title': 'app.py', 'language': 'python', 'defs': 'function alpha_helper'}
{'ax_path': 'notes.txt', 'ax_format': 'txt', ...}

and the issue's real-tree case (crates/xerj-autoindex/src, 53 .rs files) goes from 0 records live / count: 0 to 53 records live / count: 53.

Tests

• sync_executor::tests::preparation_extracts_code_from_extensionless_snapshot_blobs — the durable-path regression the issue asked for (fails before the fix).
• extract::code::tests::extracts_from_extensionless_blob_via_sniffed_logical_name — the same contract at the extractor's own boundary.

Local gate

• cargo test -p xerj-autoindex --lib: 530 passed / 0 failed
• cargo fmt --all --check: clean
• cargo clippy -p xerj-autoindex --all-targets -- -D warnings: clean
• ES-YAML conformance suite: 1366 passed · 0 failed · 3 skipped

Noted, out of scope

Other extractors (txt, html, pdf, docx) use their content path only as a fallback title when the document has no intrinsic one, so on the durable path a titleless file can still get a blob-ordinal title. Cosmetic, no document loss — worth a small follow-up that threads logical_name into those fallbacks.

Fixes #294

🤖 Generated with Claude Code