fix: new notes skip indexed ids, and fetch records only indexed notes

[fduple]

What this changes

write_note takes a taken_ids set and skips those ids as it skips an existing file; the five paths that fetch and save a note record a URL only when sync has indexed the new note under its own id.

Why

note mv keeps a note's id while renaming its file, so after note mv topic topic-moved the filename topic.md is free while the id topic is taken. write_note checked only for an existing file, so the next note titled "Topic" was written to topic.md and sync refused it (id collision: 'topic' already belongs to 'research/notes/topic-moved.md'). The fetch paths then recorded the URL (and with --save-assets the screenshot) under the moved note's id, so the duplicate-url check answered URL already fetched as note 'topic' for the new URL and the fetched text was never indexed. note new, MCP create_note and the research map-of-content reported that id for a file sync never indexes. graph stub and repair --stub: the link resolver never matches a ref by slug, so a stub for [[foo bar]] could take an id another note holds, be refused by sync, and leave the link broken.

On main, with a stub provider that titles every page "Topic":

fetch https://a.example.org/x --save-assets  -> note_id "topic"
note mv topic topic-moved
fetch https://b.example.org/y --save-assets  -> exit 0, "topic"

leaves sources (https://b.example.org/y, topic), the screenshot under topic, topic.md unindexed, and a repeat of the second fetch refused with DUPLICATE_URL naming topic.

How

Prevention: indexed_note_ids(conn) feeds write_note(taken_ids=...) at every caller that derives an id: fetch, fetch-batch, research's _save_result and map-of-content, escalation ingest, fetch_and_save, note new, MCP create_note, graph stub and repair --stub. The second fetch above becomes topic-2, the map-of-content gets a -2 id too (research now reports the id it was written under), and the stub foo-bar-2, which [[foo bar]] resolves to through its title.

Guard: after sync, the five fetch paths call discard_unindexed_note; when another file took the id between write_note and sync, the file and its raw artifact are removed and the error names the file and the id's owner. fetch and escalation ingest exit 1 with NOTE_NOT_INDEXED (the item stays claimed); fetch-batch lists the URL in failed_urls with phase index; _save_result skips it with a stderr warning; fetch_and_save raises RuntimeError.

How it was verified

• tests/test_cli/test_note_id_taken.py (15 tests) drives each case through each caller and forces the race at each fetch path (a second file indexed under the new id after write_note returns); tests/test_core/test_note.py adds one test for taken_ids. Against main all 16 fail (14 on assertions, the fetch_and_save race test because no RuntimeError is raised, the test_note.py test with a TypeError).
• Any one changed CLI or MCP file restored to main fails its tests on their assertions; core/note.py or core/fetcher.py restored alone fails on a missing import, so their behaviour was also removed by hand, and the matching tests fail, as does each race test with its guard call disabled.
• pytest tests/: 1735 passed on 3.11, 3.12, 3.13 and 3.14 (Linux). ruff check src/ tests/ is clean. Windows was not run.

Limits

• note new --template with an existing template (the default note is one) uses its own filename loop and is not covered: after the note mv above it still returns id topic for a file sync refuses.
• note new, create_note, the map-of-content and the stubs get no guard after sync; they record nothing under the id, so a race there only misreports it.
• Unchanged: note mv still leaves the moved note's sources row with note_id NULL and deletes its assets rows, --save-assets still overwrites an existing screenshot.png, and graph stub still reports the link text as each stub's id.

---

Scope check — the review is much faster when the diff contains only the change being proposed:

• No unrelated reformatting, whitespace, or prose rewriting outside the lines this change needs
• CHANGELOG.md touched only to add an entry under [Unreleased]
• No edits to user-facing output strings, docstrings, or test fixtures unrelated to this fix
• If you ran a formatter or an automated cleanup pass, its results are in a separate PR