test: cover file metadata propagation

[HarperZ9]

Summary

Adds deterministic integration coverage proving whether extractor metadata survives the complete file-ingestion path into stored chunk metadata. The currently lost extractor fields are represented as narrow strict xfails so unrelated regressions still fail and a future propagation fix becomes an XPASS failure until the markers are removed.

Linked context

Closes #77

PR Size

• Small (< 100 LOC delta)
• Medium (100-499 LOC delta)
• Large (500-999 LOC delta)
• XL (1000+ LOC delta — must have been pre-discussed in an issue)

PR Type

• Bug fix
• New feature
• Prompt change (prompt label required; eval results required below)
• Test
• Docs
• Refactor / chore

Context / Motivation

The existing integration helper called Ingestor directly, so it could not detect metadata dropped between extractor output and MemoryStore.add. Issue #77 asks for pipeline-level proof using offline text and Markdown fixtures.

Changes

• Run the existing .txt and .md fixtures through the real Anchor.ingest_file path.
• Verify source, generated questions, timezone-aware timestamp, retrieval, and stored content.
• Add strict KeyError xfails for source format, text offsets, and Markdown heading metadata currently lost before storage.
• Guard every xfail with explicit structure and content assertions so unrelated failures cannot be misclassified.

How to test

1. uv run pytest -q tests/integration/test_ingestion_pipeline.py gives 2 passed and 4 xfailed.
2. uv run pytest -m "not eval" -k "not test_malformed_python_raises_value_error" gives 152 passed, 4 xfailed, and 37 deselected. The excluded Windows path-escaping failure is pre-existing and outside this change.
3. uv run pre-commit run --files tests/integration/test_ingestion_pipeline.py passes the file-scoped formatting, lint, type, AST, secret, and hygiene hooks. The environment-wide pip-audit hook reports 17 pre-existing dependency advisories.

Prompt Change Eval Results

Not applicable.

Checklist

• I ran the relevant local checks.
• I updated documentation or confirmed no docs changes were needed.
• I linked related issues or pull requests and confirmed this is not duplicate work.
• This change stays within the stated scope of the PR.

Notes for reviewers

The xfails intentionally use both strict=True and raises=KeyError. Missing chunks, malformed stored records, wrong fixture content, invalid timestamps, or failed Markdown section selection remain ordinary failures. When metadata propagation is implemented, these tests become XPASS failures and should be converted to normal passing assertions.

AI assistance disclosure: OpenAI Codex assisted with drafting the tests and reviewing the final diff. The targeted and deterministic checks above were run locally.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 5beba004-e1cd-41fd-9a8e-d0148d3247f3

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Thanks for opening your first pull request in this repository, @HarperZ9.

Before review starts, make sure the PR checklist is complete and the fastest relevant local checks have been run. The expected workflow is documented in CONTRIBUTING.md.

[HarperZ9]

Consolidating my overlapping submissions into #102. That PR now preserves this version's source, question, timezone-aware timestamp, content, heading-content, and text offset guards, and adds the direct extractor-contract and negative-control checks. It also restricts expected failures to a dedicated missing-field exception so unrelated failures remain visible. Closing this older PR to leave one review target for #77; its branch remains available. No production changes are included.