OKF: document the log-heading deviation from §7 - #525

A sync check found the bundle already in sync, with one instructive
non-finding: the single "broken cross-link" warning points at a file that
EXISTS — docs/projects/2605-.../course-landing-components-2026-08.md. The
validator declines to resolve outside the bundle and tolerates it under §5.3.
Acting on that warning would have broken a working link.

What is real: **33 of ~80 warnings are §7 date-heading violations**, and they
come from this file's own convention. The spec's template is a bare
`## <YYYY-MM-DD>` with bullets beneath, which assumes one entry per day. This
bundle lands several thematic entries per day — 12 on 2026-08-21 alone — so
conforming would produce a dozen identical headings and destroy scannability.

Keeping `## YYYY-MM-DD - what changed`, and saying so in the header. The reason
to write it down is not the warnings, which never block: it is that a session
reading validator output otherwise either collapses the headings destructively,
or learns to ignore warnings wholesale — and the second is how a real one gets
missed. The note tells a reader which class to filter and why.

okf validate --strict conformant; bin/hugo-build clean. Log file only.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`okf_validate.py .okf --strict` **exits 1**, and has for some time. It prints
`✓ conformant (N warnings)` — true, since conformance (§9) means no ERRORS — and
I read that line as a pass every time without checking `$?`. CLAUDE.md requires
this gate before bundle commits. Verified now: without `--strict` exit 0, with
`--strict` exit 1.

That is evidence-over-assertion failing in its most ordinary form: the
human-readable summary and the machine signal disagreed, and the friendlier one
won for a whole session.

**Counts measured, not estimated.** 82 warnings: 57 §7 date headings, 23 missing
recommended fields. The first draft of this note said "33 of ~80" — an
unmeasured number inside a note whose entire purpose was separating known noise
from real signal.

Recorded alongside it, because it is the opposite error: the one "broken
cross-link" warning points at a file that EXISTS; the validator declines to
resolve outside the bundle and tolerates it under §5.3. Acting on it would have
broken a working link. One warning class was real and ignored; another looks
real and must be.

Reframed from "accepted noise" to a KNOWN-RED GATE, and tracked in
docs/projects/2608-site-design-system/README.md with the two jobs that make it
green: restructure log.md so same-day themes sit under ONE dated heading (a
whole-file restructure that conflicts with parallel sessions appending to it —
do it when the repo is quiet), and add `timestamp` to the 23 concepts missing
it, anchored to each file's last commit time, never invented.

Gates, stated precisely this time: `okf_validate.py .okf` exits 0, conformant,
81 warnings. `--strict` exits 1 by design until the above lands.
bin/hugo-build clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>