Add source_quality label and a health warning for speculation-only knowledge (#258)

[konsone]

What this PR does

The narrow first slice agreed on #258: source_quality: high | medium | speculation on raw sources, one shared reference listing the exact signals, and a Source quality health check. Capture labels, compile gates; this PR adds the label and the reader, not the gate.

• references/source-quality-check.md: the single reference. Deterministic rule, no network or model call, so a reviewer can check a label by reading the note. Lists the exact signals for medium as requested.
• references/ai-first-rules.md: the field in the type: source schema, described as orthogonal to capture_scope ([Feature] Add local source-payload completeness gates to prevent URL-only or excerpt-backed knowledge #194): one says how much was kept, this says whether it was worth keeping.
• scripts/vault_health.py: check_source_quality warns when a knowledge note cites only speculation sources. A source with no field is unjudged and never reported, so a vault written before the field stays silent. Daily, log, index and other activity notes are skipped.
• source_quality_policy: warn | silent in .vault-config.json, same never-inferred contract as rewrite_policy ([Feature] Make the /obsidian-ingest confirm-before-rewrite step (#239) a vault-level policy: confirm (default) or unattended #250). No value blocks a capture.
• The link-indexing block of check_source_payload moved into a shared _dependents_of helper so both checks use one implementation.

The field is meant as the contract any later verifier (SIFT, a claim-level check, a human) writes to; consumers read only the field.

Related issues

Related to #258 (not closing: the ingest step 6 behaviour and writing the field from capture commands follow separately, as discussed there).

How I tested it

• 17 new tests in tests/test_source_quality.py, synthetic fixtures only: all-speculation warns; one better or unlabelled source keeps it quiet; unrecognised value ignored; independent of capture_scope; source citing a source and uncited sources do not ring; daily, log and index notes do not ring; template-comment values parse; policy defaults and silent.
• Full suite on current main: 1011 passed; 7 failures in test_groq.py / test_podcast_resolution.py are identical on an untouched main.
• Ran against a copy of a real 848-note vault with labels written by a local pilot writer (not in this PR): 0 findings without labels, 3 with, all notes resting on a single derived Q&A or agent-report source. No false positives on daily/log/index notes.

Honest limits

• Qualitative check only: no gold-labelled hit rate or false-positive rate.
• Only type: source notes cited by wikilink are covered. Transitive propagation (concept to synthesis) and non-source capture types are follow-ups.
• Nothing writes the field yet, so on an existing vault the warning stays silent until the capture commands do.
• One divergence to confirm: the rule requires a traceable locator for medium. An author or date with no locator does not reach medium on its own.

AI-first rule compliance

• Follows references/ai-first-rules.md; no command that writes to the vault changed
• No banned substitution characters in instruction files

Checklist

• Searched existing issues and PRs
• CHANGELOG.md updated under Unreleased

🤖 Generated with Claude Code

[eugeniughelbur]

@konsone thanks, this is the narrow slice we agreed on in #258, and it's careful about the right things.

What I checked:

• All 17 new tests fail on current main, and CI is green on Linux and Windows.
• _dependents_of is a line-for-line lift of the old block in check_source_payload, so the payload check behaves the same.
• An unlabelled source keeps a note quiet. That makes the check safe on every existing vault, and there's no policy value that blocks a capture.
• The activity-note skip list covers the daily, log and index cases that would otherwise make people mute it.

One small edge, not a blocker. _fm_field strips quotes before your # split, so source_quality: "speculation" # high | medium | speculation comes out as speculation" and is ignored. It fails quiet, so nothing gets flagged wrongly. But the template comment plus our "double-quote when in doubt" convention makes that line likely. Splitting off the comment first, then stripping quotes, fixes it. A test with that exact line would pin it.

On your open divergence: yes, medium needs a traceable locator. An author or date with no way to find the source again stays speculation.

[konsone]

Thanks for the careful read, and for settling the medium locator question.

Fixed the quote edge in the follow-up commit: the # comment is now split off first, then the quotes are stripped, then the value is lowercased. Pinned with your exact line, source_quality: "speculation" # high | medium | speculation, plus single-quoted and upper-case variants. The related tests pass locally (68 across test_source_quality.py, test_source_payload.py and test_vault_health_precision.py); CI will rerun on the push.

Nothing else changed in the slice. Writing the field from the capture commands and the step 6 behaviour I will leave for a separate PR, once you are happy with the field's shape here.

[eugeniughelbur]

@konsone merged in fe84aa9. Thanks for the quick fix on the quote edge and for pinning it with that exact line.

#258 stays open for the next slices you named: writing the field from the capture commands, and the ingest step 6 behaviour.