fix(mcp): fix small note-tool defects from the v0.24.0 live run

[phernandez]

Fixes #1634.

Summary

1. Out-of-range line read

• slice_note_content (markdown/sections.py) returns a NoteSliceError when a line range starts after the last line. The message is start_line 500 is past the end of the document (17 lines). Before this change it clamped only the end, which produced the inverted range "Lines 500-17".
• The API turns this error into a 404 with that message as the detail, the same way it handles other slice failures. read_note (text and JSON), cat with max_tokens, and the CLI all report the message as an error.
• cat slices plain line ranges on the client side, so that path raises the same message.
• Line 1 of an empty document can still be read.

2. Markdown links that climb out of the project

• New helper climbs_out_of_project(target, source_path) in markdown/path_links.py. It is true for a path target where resolve_project_path returns None.
• Relation rows are built against the note's own path at four points: accepted write, move, the batch indexer, and the EntityService graph publish. All four now drop these targets instead of storing an unresolved links_to row that can never resolve.
• Canonical Markdown is unchanged. Relations are derived state.
• The rule applies to every path-shaped target, including wikilinks spelled [[../../x.md]]. feat(core): index Markdown links to project files #1514 already resolves those with the same rule.

3. write_note conflict names the wrong note after a move

• The service outcome AlreadyExists now carries the NoteLocation of the note found at the path. On a lost create race (a 409 from create), it looks up the note that won the path.
• NoteAlreadyExists in schemas/v2/note_write.py gains external_id: str | None = None and permalink: str | None = None. The change is additive: no field is renamed or removed, and a payload with only file_path still validates.
• The tool now reports the stored permalink of the note at the path. In a workspace context it qualifies that permalink the same way the success path does. The JSON payload adds external_id and sets file_path.
• If the server cannot name the note (for example, an older server), the suggested edit_note/read_note identifier falls back to the file path, not the permalink computed from the request.

4. recent_activity

• The tool description, docstring, and prompt no longer say "all projects" for a bare call. A bare call uses the active project, then the default project. Discovery across projects happens only when neither resolves. The prompt header now reads "the default project".
• When a tool returns an empty list, FastMCP sends no content blocks at all, so a text-only client got nothing. A new EmptyListResultMiddleware (mcp/empty_results.py) adds a single [] text block in that case: the same JSON shape a non-empty list renders as, so clients that parse the text (joined or not) still get valid JSON. Structured content stays {"result": []}. The middleware applies to every tool, so tail gets the same fix.
• The server instructions now say find needs project when no project-qualified path is given, alongside grep and tail.

5. Local authorship keys

Already resolved on main by 91d79bd. This PR does not touch it.

Behavior changes worth noting

• Two existing tests pinned the old behavior and are updated:
  • read_note(start_line=1000) used to return empty content. It now raises the past-end error.
  • The MCP conflict payload used to have file_path: None. It now holds the existing note's path.
• The already_exists HTTP payload now includes external_id and permalink. The cloud consumes this schema, and the addition is backward compatible.

Verification

• just fast-check: lint, format, and typecheck all clean.
• BASIC_MEMORY_TESTMON_SELECT_FLAGS="--import-mode=importlib" just fast-test <touched test paths>: passed.
• just test-sqlite:
  • unit: 7962 passed, 58 skipped
  • integration: 688 passed, 15 skipped
• just man-regen: all pages already match their source.

🤖 Generated with Claude Code

https://claude.ai/code/session_01APFUk2bjEwMptMQqpRhjea

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-09T03:31:56.629248Z	04d4684	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f65fceaf33

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".