Fix obsidian-cli eval timeout causing false smoke-test failures

[yolo-sam]

Fixes #17

Problem

test/test.sh reported ~18 false failures: assertion blocks returned empty strings, so checks compared '' against expected values and failed.

Root cause

obsidian-cli eval only returns a value if its async IIFE settles within the same microtask flush. Any setTimeout — confirmed empty down to 50ms — yields to the event loop and the eval times out to empty. Microtask-only async and blob-URL fetch reads resolve fine inside one eval, which is why the fast blocks already passed.

Fix

The six slow blocks (THEME_RERENDER, DEDUPED_NAVIGATION, DEFERRED_STATE_NAVIGATION, MARKDOWN_EMBED, SIZED_MARKDOWN_EMBED, CANVAS_EMBED) are restructured to trigger-eval → bash sleep → read-eval. State that must survive the wait (theme classes; the deferred-leaf getViewState monkey-patch + its original) is stashed on window. No assertion intent or expected value changed.

cleanup() had the same latent bug — its in-eval 300ms setTimeout returned empty, silently skipping the leftover-leaf verification. Split into detach → bash sleep → inspect/restore so cleanup verifies again.

Header comment updated npm test → bun run test. CI workflows untouched.

Validation

bun run test against the live vault, stable across two consecutive runs:

• Outer + embed assertions: 35/35 pass
• Inner (iframe postMessage): 18 passed · 0 failed · 0 skipped
• All checks passed, exit 0.

[yolo-sam]

Adversarial review — Round 1 (Claude + Codex), head 1e2e313

Claude reviewer: VERDICT MERGEABLE. Two LOW advisory notes (no fix required): cleanup() dropped detachErrors from its JSON; deferred monkey-patch window slightly wider.
Codex reviewer: VERDICT NOT MERGEABLE, 3 findings — all accepted and fixed in 6205ff1.

Resolutions

1. [HIGH] MARKDOWN_EMBED read eval re-found the wrong leaf (test/test.sh:418). DEDUPED_NAVIGATION and DEFERRED_STATE_NAVIGATION already open _html-docs-test-embed.md in earlier leaves, so re-finding 'first leaf by path' could read a non-preview leaf → false failure. Fixed (6205ff1): trigger eval stashes the exact opened leaf on window.__hvMarkdownEmbedLeaf; read eval reads that leaf and deletes the stash. (SIZED uses a unique note, CANVAS a unique file — unaffected.)
2. [MED] Deferred getViewState monkey-patch could leak (test/test.sh:350). The original try/finally restore now spans a bash boundary; a throw before the restore eval would leave the live leaf patched. Fixed (6205ff1): cleanup() (the trap EXIT path) defensively restores getViewState from window.__hvDeferredState if present.
3. [MED] Theme classes could be left flipped (test/test.sh:253). Same split-finally issue — a throw mid-flow would leave Obsidian's theme toggled. Fixed (6205ff1): cleanup() defensively restores the original theme classes from window.__hvThemeState if present.

Pushback

• Claude's detachErrors note: keeping as-is. No assertion ever consumed detachErrors; a leaf that fails to detach still surfaces via the remaining array → exit 1. Net diagnostic loss is a less-specific message, not a missed failure.

Validation at new head

bun run test against the live vault: all outer + embed assertions pass, inner (iframe) 18 passed · 0 failed · 0 skipped, exit 0.

Confirm round (both families) running at 6205ff1.

[yolo-sam]

Adversarial review — Round 2 (confirm), head 6205ff1

Consensus reached. Fresh reviewers from both families re-read the full diff and verified all three round-1 resolutions:

• Claude: all 3 CONFIRMED RESOLVED, no new issues — VERDICT MERGEABLE
• Codex: all 3 CONFIRMED RESOLVED, no new issues — VERDICT MERGEABLE

Both confirmed: MARKDOWN_EMBED now reads the exact stashed leaf (no path re-find); cleanup() restores the deferred getViewState patch and theme classes on the guaranteed trap EXIT path, before detach, and is idempotent with the success path (no double-restore). No in-eval setTimeout remains; bash -n clean.

Flipping out of draft and assigning @smcllns. Not merging — Sam merges.