fix(lint): preserve Obsidian heading, block, and embed link syntax

[Aldominguez12]

Problem

The wikilink regex in openkb/lint.py captured the whole [[...]] body as a single page target. Valid Obsidian syntax therefore never matched a known target:

• [[page#Heading]] / [[page#^block]] — heading and block links
• ![[file.png]] / [[report.pdf]] — attachment embeds and links
• [[#Heading]] — same-page fragment links
• ![[concepts/x]] — note embeds

Consequences:

• openkb lint --fix destroyed valid links: strip_ghost_wikilinks demoted them to plain text. A wiki-wide sweep silently mangled hand-written notes in explorations/ — a data-loss path.
• openkb lint reported them all as broken links (false positives).
• Note embeds created no graph edges in visualize and could produce false orphans.

Fix

• Parse the embed marker, target, fragment, and alias as named regex groups.
• Validate only the page target; fragments survive fuzzy canonical rewrites ([[concepts/Gist_Memory#Notes]] → [[concepts/gist-memory#Notes]]).
• Pass through attachment embeds/links (extension whitelist) and same-page [[#Heading]] links untouched.
• Treat note embeds ![[concepts/x]] as regular page links: validated, rewritten keeping the !, counted as incoming links for orphan detection, and picked up as graph edges by visualize (which reuses _extract_wikilinks).

Verification

• New TestObsidianSyntax coverage: heading/block/fragment preservation through direct and fuzzy matches, ghost demotion without ! residue, attachment/note embed handling, and a regression test asserting a hand-written explorations/ note survives a wiki-wide lint --fix byte-for-byte.
• pytest tests/test_lint.py tests/test_lint_cli.py tests/test_visualize.py tests/test_chat_session.py tests/test_remove.py green; ruff check/format and mypy openkb/lint.py clean.

🤖 Generated with Claude Code

[KylinMountain]

Nice fix — this closes a real data-loss path and the test coverage is thorough.

One thing worth addressing before merge: _WIKILINK_RE backtracks O(n²) on long runs of unmatched [, and the new optional frag/alias groups make it ~10× slower than the old regex — a doc with [×20000 takes ~12s per scan (and lint scans each file twice), so a malformed/pathological doc (ASCII art, code fence, base64, stray LLM output) can hang the linter for seconds.

Excluding [ from the target class — (?P<target>[^\[\]|#]*) — makes it linear again, and as a bonus fixes [[[a]]] capturing [a as the target. The other edge cases I noticed are minor.

[Aldominguez12]

Good catch — confirmed the quadratic behavior locally (20k unmatched [ took ~9.5s per scan, linear after the change) and that [[[a]]] now matches the inner [[a]] instead of capturing [a as the target. Fixed in e03e322 by excluding [ from the target class, with regression tests for both the pathological input and the nested-bracket case. All lint tests pass (88/88), ruff and mypy clean.

[KylinMountain]

Small follow-up on the same regex: e03e322 fixed the [-run case by excluding [ from target, but the two new groups still include [ — frag is #[^\]|]* and alias is [^\]]+ — so the same O(n²) backtracking survives; it just moves to unclosed [[…#… / [[…|… runs.

Repro against the PR head, calling the real strip_ghost_wikilinks + _extract_wikilinks (lint scans each file twice):

• "[[a#" * 10000 (a long run with no ]) hangs for seconds; doubling the count ~4×s the time.
• The trigger is narrow, though: put a single ] roughly every ~10 links and it drops back to ~20ms. So it only bites on a long ]-free region densely seeded with [[…#/[[…| — corrupted/truncated or adversarial input, not normal notes (a 175KB note full of closed [[a#b|c]] links is ~9ms). Low severity — just the same class e03e322 set out to close.

One-liner — extend the same [-exclusion to both inner classes:

r"(?P<embed>!)?\[\[(?P<target>[^\[\]|#]*)(?P<frag>#[^\[\]|]*)?(?:\|(?P<alias>[^\[\]]+))?\]\]"

(#[^\]|]* → #[^\[\]|]*, and [^\]]+ → [^\[\]]+)

Verified against the real functions: every pathological shape goes linear ([[a#×10000 → ~15ms), and matching on valid Obsidian syntax is byte-for-byte unchanged — heading/block/fragment rewrites, note/attachment embeds, [[#Heading]], aliases, and ghost demotion all identical. Might be worth a regression test on the frag/alias shapes alongside the existing [-run one.

[Aldominguez12]

Thanks for the thorough follow-up — you're exactly right. e03e322 only bounded target; frag (#[^\]|]*) and alias ([^\]]+) still admitted [, so the same O(n²) just moved to [[…#… / [[…|… runs.

Reproduced against the real strip_ghost_wikilinks + _extract_wikilinks: [[a#×10000 ran ~15.5s per pass (≈4× on doubling — quadratic), and your one-liner takes it to ~7ms (linear). Applied it verbatim in 3ee0da7:

• frag: #[^\]|]* → #[^\[\]|]*
• alias: [^\]]+ → [^\[\]]+
• widened the class comment to cover all three inner classes, not just target
• added the two regression tests you suggested (test_unmatched_fragment_run_scans_in_linear_time / …_alias_…) alongside the existing [-run one — full lint suite 90/90.

Matching on valid Obsidian syntax is unchanged across the corpus (headings, block refs, aliases, note/attachment embeds, [[#Heading]], ghost demotion). The only shapes that stop matching are fragments/aliases containing an unmatched [ — malformed input, and the same trade-off already accepted for target. Thanks again!