lint: .raw/ archive is walked as link-resolution candidates, inflating ambiguous-target findings

[JuliusGruber]

Summary

lint walks .raw/ and indexes archived captures as link-resolution candidates. Obsidian does not — it ignores every dot-prefixed folder. Because a capture and its derived source page share a basename by design, this makes nearly every [[Source Title]] link in a healthy vault report as ambiguous-target.

On my vault the effect is 307 reported findings where 21 are real. The vault is fine; the signal is buried.

Root cause

claude_obsidian/lint_engine.py:56 skips a fixed set of directory names rather than applying Obsidian's dot-prefix rule:

_IGNORED_WALK_DIRS = {
    ".git",
    ".obsidian",
    ".vault-meta",
    ".cache",
    ".pytest_cache",
    "__pycache__",
    "node_modules",
}

.raw is absent, so _walk_files() descends into it. This is inconsistent with the package's own design intent, which treats the archive as hidden — skills/wiki-ingest/SKILL.md:9 describes "inbox/ as visible staging and .raw/ as the legacy immutable archive", and the dot prefix is what makes it invisible in the app.

Reproduction

Any vault where .raw/<Folder>/<Title>.md was ingested into wiki/sources/<Title>.md — i.e. the normal ingest outcome:

$ python scripts/claude-obsidian.py lint --vault "$VAULT" --format markdown

- Pages scanned: 153
- Links scanned: 2892
- Issues found: 307

## Ambiguous Targets
- `wiki/concepts/Agent Readability.md`:108 -> `Effective Harnesses for Long-Running Agents`;
  candidates: `.raw/HarnessEngineering/Effective harnesses for long-running agents.md`,
              `wiki/sources/Effective Harnesses for Long-Running Agents.md`

The second candidate is the only one Obsidian can actually resolve to.

Impact

Measured on a 153-page vault, comparing current main behaviour against the same vault with .raw added to the skip set:

Category	Reported	Real	Inflation
ambiguous_targets	253	6	247
stale_index_entries	28	2	26
orphans	14	1	13
dead_links	9	9	—
empty_sections	3	3	—
Total	307	21	286

The orphan inflation is a second-order effect worth calling out: a wiki/sources/* page is linked from wiki/index.md, but because that inbound link resolves ambiguously it isn't counted as inbound, so a correctly-linked page is reported as an orphan. The three categories fail together.

This also makes --strict unusable in automation, since a healthy vault never reaches zero findings.

Suggested fix

Add .raw to _IGNORED_WALK_DIRS. I ran this against the vault above: findings drop 307 → 21, and every remaining finding is genuine.

Two things I checked before proposing it:

• Provenance validation is unaffected. _provenance_errors() (lint_engine.py:823, called at :1084) reads the ledger from the vault root directly rather than from _walk_files() output, so skipping the directory does not skip the manifest audit. provenance_errors stayed at 0 after the change.
• pages_scanned and links_scanned are unchanged (153 / 2892), confirming .raw/ files were only ever acting as link targets, never counted as pages.

A more general fix would be to skip any dot-prefixed directory, matching Obsidian's actual rule and covering user-created hidden folders too. That is a wider behaviour change, so I've suggested the narrow version — happy to send a PR for whichever you prefer.

Environment

• claude-obsidian 2.1.0, engine 1.1.1, report schema 1
• Python 3.13.3, Windows 11
• Vault: 153 pages, .raw/ archive of 24 markdown captures + 2 PDFs

[JuliusGruber]

Working through the rest of my report surfaced two more divergences with the same root cause: Obsidian appends .md to an extensionless wikilink before trying an exact match, and the engine doesn't. Both produce false positives on a healthy vault, so I'm adding them here rather than opening separate issues.

2. .canvas / .base are treated as stem-addressable

_EXTENSIONLESS_WIKILINK_SUFFIXES (lint_engine.py:54) includes .canvas and .base:

_EXTENSIONLESS_WIKILINK_SUFFIXES = frozenset({".md", ".canvas", ".base"})

The comment says these are "addressable by stem in the app". In practice [[dashboard]] resolves to dashboard.md whenever it exists — the .base is only reachable as [[dashboard.base]]. So when a .md landing note and a same-named .canvas/.base coexist, the engine reports ambiguity where Obsidian has a deterministic answer.

This is a common layout, because it's the workaround for canvas and base files not being stem-addressable in the first place:

<!-- wiki/meta/dashboard.md -->
The interactive dashboard lives in [[dashboard.base]]. Open it directly, or use the embed below.
![[dashboard.base]]

My vault adopted exactly that after weighing the alternative (rewriting call sites as [[Wiki Map.canvas]]) and rejecting it, so the six findings are reporting the intended design as a defect.

Suggested fix: when candidates differ only by extension and one is .md, resolve to the .md instead of reporting ambiguous-target.

3. A page named <Name>.md.md is reported as an orphan

Given wiki/concepts/CLAUDE.md.md and 28 pages linking [[CLAUDE.md]], Obsidian tries CLAUDE.md.md first and lands on the concept page. The engine matches the vault-root CLAUDE.md exactly, so those 28 inbound links bind elsewhere and the concept page reports as having none.

The <Name>.md.md filename is deliberate — it's how you make [[CLAUDE.md]] reach a page about CLAUDE.md, and my vault verified it resolves correctly across all referencing files before adopting it. The same trick appears with Getting Claude to Actually Read Your CLAUDE.md, where the title genuinely ends in .md.

Suggested fix: resolve [[L]] against L.md before falling back to an exact match on L, matching Obsidian's order.

Combined effect

On my 153-page vault, all three divergences together:

	findings
Reported by current main	307
Divergence #1 (.raw/ walked)	286
Divergence #2 (.canvas/.base)	8
Divergence #3 (.md.md orphan)	1
Genuine defects	12

The 12 real ones were 9 dead links and 3 malformed empty sections; I've fixed those and the vault now lints clean under a local wrapper implementing all three rules. Verified by replaying the original report through it — it yields exactly the 12.

Happy to send a PR covering all three if you'd like; #2 and #3 are both small changes in the resolver, and #1 is the one-line skip-set addition.