[dbccpage]

🤖 Generated with Claude Code

This is a follow-up to #45, from the third external review of that PR (at 188213c). Two of its four findings survived the merge. Each reproduces on main, and the new tests fail there.

#	Finding	On main	Fix here
P1	A write from another SQLite connection is invisible to total_changes, so a 0.2 committed elsewhere beside an admitted 0.1 summed to Decimal('0.3')	Reproduced	The audit also compares PRAGMA data_version, which moves on any other connection's commit. Exact-decimal eligibility is revoked for good, because the lost source types cannot be recovered, only refused
P1	set_watermark() synced the trusted counter without auditing first, so it absorbed an unobserved write. create_and_seed_table() did the same	Reproduced	Both public calls audit before they sync. apply_write_event syncs its own write, then writes its watermark through an internal path that does not audit
P1	Cold MIN/MAX converted floats through Decimal(str())	Fixed by #45's fourth commit	One gap closed: SortedMultisetExtremumCache.uncacheable_reason is now set on the UNSUPPORTED outcome, as for AggregateCache
P2	An unregistered float GROUP BY escaped as a bare TypeError	Fixed by #45's fourth commit	Test added; it passes on main

Ownership contract. The adapter owns the substrate's writes.

• SQLite: a write the adapter did not make is detected, from this connection or any other, and revokes decimal certification.
• DuckDB: it has neither counter, so the guarantee covers adapter writes only. The substrate says so in its own comments.

Evidence (local, Python 3.13)

• Full suite: .venv/bin/python -m pytest -q -p no:cacheprovider --ignore=tests/production_native.py exits 0.
• Types and formatting: mypy smplcache reports no issues, and black is clean.
• New tests: 8 tests in tests/test_source_provenance.py.
  • On main, 4 fail: the other connection, the watermark, the seed and the extremum mark.
  • On the reviewed commit 188213c, 7 of the 8 fail. The eighth, watermarked adapter writes staying exact, is a positive control.

Branch note. This branch reused #45's name after the merge. Its diff against main is only these 3 files, and the extra commits in the list are #45's pre-squash history.

https://claude.ai/code/session_0194kAmYv4pLMyrspzKEWExZ

---

Generated by Claude Code

[dbccpage]

Live Database Containers is red, but the failure is not from this PR.

• The failure: 1 failed, 56 passed. The failing test is tests/production_native.py::test_production_read_contention_during_publication[bounded], with "native source did not catch up within 30 seconds".
• It is timing-flaky on main. It failed on 49a3a5c and d52e174 and on every Admit floats as values; refuse fake exactness as float_not_exact #45 head, and passed on 9e55b2f and 3a838a2.
• Cause: the bounded worker drains about 13 generations/s against the writer's 20 commits/s, so the backlog takes 29 to 41 s against a 30 s limit. This PR touches only the cold-rebuild substrate (substrate.py, normal_forms.py), not the production worker.
• Fix: none exists yet. It changes the publication protocol and waits on the maintainer's choice.
• Not re-run: the repository's Actions minutes are limited.

---

Generated by Claude Code