fix(core): project-scoped forget, dim-safe cosine, precise question detection

[krishpinto]

What

Three correctness fixes surfaced while auditing the engine (grasp-core). Each is independent, small, and unit-tested.

1. forget() over-matched sibling projects

The processed_files cleanup used LIKE '%slug%' — an unanchored substring match. Consequences:

• Forgetting c--projects-grasp also cleared records for c--projects-grasp-core (substring), forcing a needless full re-import of the sibling.
• _ / % in a slug acted as SQL wildcards.

Now it matches the project's transcript directory as an anchored, escaped prefix. A new escape_like helper escapes %, _, and \ so Windows paths stay literal under LIKE … ESCAPE '\'.

2. cosine() silently mis-scored mismatched dimensions

a.iter().zip(b) truncates to the shorter slice, so a stale embedding of a different length produced a bogus (partial) similarity. The embeddings table already stores a per-row dim, so a future model/DIM change can leave differently-sized vectors behind — those must not corrupt ranking or semantic edges. Now returns 0.0 on a dimension mismatch.

3. is_question() false-positives on substrings

A bare contains("what ") flagged "somewhat slow" as a question. Now interrogatives match only at a word boundary (start of string or after a non-alphanumeric char).

Tests

• escape_like_escapes_wildcards_and_backslashes
• cosine_of_mismatched_dimensions_is_zero, cosine_of_identical_unit_vectors_is_one
• question_detection_respects_word_boundaries

Full grasp-core suite green (29 passed, 2 pre-existing ignored), and grasp-cli builds clean.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Fixed similarity scoring so vectors of different lengths no longer yield misleading results.
  • Improved question detection by requiring interrogative keywords to appear at real word boundaries, reducing false matches inside other words.
  • Strengthened project “forget” cleanup to precisely target only records under the project’s transcript directory prefix, avoiding accidental deletion of unrelated items.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 8d2479bc-6830-4b42-90a4-21a812cb2dc7

📥 Commits

Reviewing files that changed from the base of the PR and between 5b20c97 and 646358b.

📒 Files selected for processing (1)

• crates/grasp-core/src/lib.rs

🚧 Files skipped from review as they are similar to previous changes (1)

• crates/grasp-core/src/lib.rs

---

📝 Walkthrough

Walkthrough

Three correctness fixes in grasp-core: cosine now returns 0.0 when input slice lengths differ; is_question switches from substring matching to word-boundary detection via a new helper; and Grasp::forget uses a new escape_like function to safely build an anchored SQL LIKE prefix for processed_files deletion.

Changes

Correctness Fixes

Layer / File(s)	Summary
Cosine similarity dimension guard crates/grasp-core/src/embed.rs	cosine returns 0.0 when slice lengths differ instead of silently truncating; test added for mismatch and identical-vector cases.
Word-boundary keyword detection crates/grasp-core/src/extractor.rs	is_question uses new contains_at_word_boundary helper instead of str::contains; test asserts true positives and rejects embedded substrings like "somewhat".
escape_like helper and anchored forget deletion crates/grasp-core/src/lib.rs	New escape_like escapes %, _, and backslashes for SQL LIKE; Grasp::forget now builds a project-directory prefix and deletes only matching processed_files rows using ESCAPE '\\', with unit tests for the helper.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐇 A cosine that slipped on unequal feet,
A keyword that hid in words incomplete,
A wildcard that wandered where it shouldn't roam—
The rabbit patched each, and brought the code home.
Now tests stand guard so bugs dare not repeat! ✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the three core correctness fixes in grasp-core and matches the changeset.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch worktree-code-review-fixes

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/grasp-core/src/lib.rs`:
- Around line 185-195: The `processed_files` cleanup is rebuilding the deletion
prefix from `config.claude_projects_dir`, which misses imports created from a
custom project path. Update the cleanup logic in this import flow to use the
actual stored project/file path that was inserted into
`processed_files.file_path` (the same `file.to_string_lossy()`-based value)
instead of reconstructing it from the configured projects dir. Keep the existing
`LIKE`/`escape_like` approach, but base the prefix on the real path associated
with the imported project so stale rows are removed correctly.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 45ef2a02-c9a9-4112-bd0c-8632f987d06c

📥 Commits

Reviewing files that changed from the base of the PR and between 28ab11a and 5b20c97.

📒 Files selected for processing (3)

• crates/grasp-core/src/embed.rs
• crates/grasp-core/src/extractor.rs
• crates/grasp-core/src/lib.rs