feat(sitegen): group thousands with U+202F, and route the last renderer through numbers

[P0w3r223]

Why

0007 §5 clause 8 of the portfolio page specification asks every grouped figure on a
published page for a narrow no-break space, U+202F. This repository writes three figures, in
three spellings from three places — which is the drift sitegen/numbers.py exists to end,
reproduced inside the repository that argues against it.

figure	written by	carried
10,000	numbers.integer()	comma
1 000-resample	a literal in sitegen/page.py:219, and a hand-typed twin in README.md	plain space
14,745	examples/validation_table.py:130, a format string of its own	comma

The third one is the point

sitegen/numbers.py opens with "no renderer is allowed a format string of its own", and
docs/decisions/0007-the-page-is-generated.md:19-22 tabulates this exact script as the
divergence the package argues against. It still had {:,}. It now calls numbers.integer(),
and the figure it writes — n = 14 745/arm — reaches docs/data/findings.json,
docs/index.html and README.md verbatim.

The separator is written "\u202f" in Python sources rather than as the character. U+0020 and
U+202F are one string in a diff, a terminal and a grep, and a module whose whole purpose is
that surfaces cannot disagree about a number should not rely on a reader spotting an invisible
one. README.md gets the character, being Markdown.

What the re-record turned up

--record moved two fields beyond the scenario string:

• recorded_on: 2026-08-21 → 2026-09-08.
• ab_lab_version: 0.3.0.dev0 → 0.4.1. The published provenance named a version two
  minor releases behind the package.

Every one of the five measured rates reproduced to the digit — 0.0538, 0.0497, 0.8077,
0.7929, 0.0110 — so the seed is honest and the version gap moved no result. That is the check
worth having, and it is why the re-record is reported here rather than assumed.

What tests/test_record.py:86 actually guards

It re-derives the sample size closed-form and asserts the recorded scenario names it. It is
re-pinned to the new spelling, and it applies the substitution itself rather than calling
numbers.integer — a guard that formats through the function it guards asserts nothing about
the formatting.

Mutation-tested, and the result corrects a natural assumption about it:

mutation	result
validation_table.py:130 back to {:,}, without re-recording	passed — the guard reads the committed artifact, not the script
the scenario string in docs/data/findings.json back to a comma	FAILED, as named
validation_table.py:130 back to {:,} and re-recorded	FAILED, as named
numbers.integer back to a comma	FAILED — all three of test_site_committed.py's byte guards

So the guard is live over the recorded evidence and over the script through a re-record, and
is deliberately blind to an un-recorded script edit. That is correct — findings.json is a
committed artifact — but it is not what the test's name suggests, so it is written down.

Verification

• pytest — full suite green.
• python examples/validation_table.py --record, then python -m sitegen.build, which
  rewrote docs/index.html and the generated: regions of README.md.
• The index checker now reads this surface clear: ok 8 separator U+202F 3.