fix(compile): drop same-batch duplicate page titles as collisions

[nickmopen]

Fixes #439

What changed

compile_kb()'s phase-1 draft loop built taken_names once before
validating any draft and never grew it as drafts were accepted, unlike
phase 2 (wikilink resolution) just below it, which already folds
survivors back in incrementally. So two same-titled (or
same-slugifying) page drafts returned in a single LLM response both
passed _draft_problem() and got filed as separate PAGE proposals
sharing id = _slugify(title).

proposals.approve() deliberately exempts PAGE proposals from the
generic "no existing artifact" guard to support the vault-edit flow, so
approving the first duplicate proposal creates the page and approving
the second finds it already exists and silently routes through
update_page() — overwriting the first approved page's content with
no error, no diff, and no "already exists" signal to the reviewer.

The fix folds each survivor's title and slug into taken_names as soon
as it's accepted, so a later draft in the same batch colliding with it
is dropped the same way it would be if it collided with a pre-existing
page or a pending proposal.

Why

Two compiled drafts landing on the same topic in one response is a
plausible LLM output (the compile prompt only tells the model to dedupe
against existing pages/pending proposals, not against its own output),
and the collision guard's own comment in compile.py shows this was a
known invariant just not enforced for the same-batch case. Silent
overwrite-on-approve with no audit signal is exactly the failure mode
the collision guard exists to prevent.

What I tested

• Added test_collision_within_same_batch_dropped to
  tests/test_compile.py, mirroring the existing
  test_collision_with_existing_page_dropped /
  test_collision_with_pending_proposal_dropped style: two same-titled
  drafts in one compile_kb() call now yield exactly one proposal, and
  approving it preserves the first draft's body.
• python -m pytest tests/ -q --ignore=tests/embeddings — full suite green
  other than pre-existing, environment-only failures unrelated to this
  change (Windows path-separator/os.getuid/symlink-privilege
  differences and a few tests that spin up a real loopback HTTP server,
  which hang under this sandbox's network policy rather than failing —
  none touch compile.py, proposals.py, or tests/test_compile.py).
• python -m mypy src — clean (pre-existing os.getuid/jinja2
  Windows-only noise aside).
• python -m ruff check src tests — clean.

CHANGELOG

Added an entry under ## [Unreleased] / ### Fixed.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: be6eac83-2c74-48b2-9a51-6424363cbaf7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[nickmopen]

closing — duplicate of #440, which already fixes this (opened earlier from a separate session under this same account). apologies for the noise.