ci: resolve the floors wheel-only; fix the 3.13 pydantic floor

[lesnik512]

Follow-up to #188.

The gap

uv sync --resolution lowest-direct falls back to building a source
distribution when a package has no wheel for the running interpreter. So the
lowest leg answers "do the floors resolve and pass on a machine with a
compiler", which is not the question. A user installing into a slim image has
no toolchain, and for them a floor with no wheel is a hard install failure that
CI reports as green.

uv sync spells the fix --no-build ("don't build source distributions"). The
leg already passes --no-install-project, so this constrains dependencies only.

What it found

pydantic>=2.8; python_version == '3.13' is a floor that exists for wheel
coverage and does not achieve it. Against main, wheel-only:

error: Distribution `pydantic-core==2.20.0` can't be installed because it is
marked as `--no-build` but has no binary distribution

Bisected on a verified 3.13.15 interpreter:

pydantic	pydantic-core	cp313 wheel
2.8.0	2.20.0	no
2.8.1	2.20.1	yes
2.8.2	2.20.1	yes

So the floor moves to >=2.8.1, and the comment above it, which asserted 2.8,
is corrected to match what the wheels actually show.

The matrix gains 3.12

#188 omitted it on the grounds that its floors resolve to the same versions as
3.11's. That is true and it is not sufficient: wheel coverage is per-interpreter,
so two legs resolving identical versions can still disagree on whether those
versions install. 3.12 passes today, but nothing was checking. The comment
justifying the omission is replaced rather than patched.

Verified

All four legs, wheel-only, with the corrected floor:

leg	resolve	pydantic	suite
3.11	OK	2.0.2	517 passed, 107 skipped
3.12	OK	2.0.2	517 passed, 107 skipped
3.13	OK	2.8.1	517 passed, 107 skipped
3.14	OK	2.12.0	517 passed, 107 skipped

The 107 skips are the Postgres tests, which run in CI. just lint-ci clean.

Credit

The failure mode came from a parallel session working on compose2pod, which
measured the same thing for PyYAML: 6.0.1 and 6.0.2 install by compiling
the sdist, so a floor of >=6.0.1 passes CI and still fails in a minimal image,
while 6.0.3 is the first with a cp314 wheel. Same class of bug, same fix.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	989	1.000	6.78	914	242	5000	5000	10000
consumer/w1/b100	1016	1.000	6.82	920	243	5000	5000	10000
consumer/w2/b10	1177	1.000	6.77	942	242	5000	5000	10000
consumer/w2/b100	1371	1.000	6.73	955	243	5000	5000	10000
consumer/w4/b10	1035	1.000	6.82	996	263	5000	5000	10000
consumer/w4/b100	1516	1.000	6.86	1035	245	5000	5000	10000
consumer/w1/b100/tfbs100	5099	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2106	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).