docs: record the 48 → 28 burn-down and correct the #393 attribution - #403

PR #402 closed the handle-leak cluster: the leak signature on the windows
leg went 18 -> 0, with zero panics and zero [build failed] packages,
measured uncapped against head 27d23e8's check-runs. The campaign
burn-down is now 85 -> 57 -> 48 -> 50 -> 48 -> 28.

The correction matters more than the number. This campaign predicted -18
from the #393 ruling; the framework change alone delivered -3, and the
remaining -17 came from closing fifteen test runtime environments that
were never closed. The production leaks #393 diagnosed were real and are
fixed -- lore's deploy loop ran N executors off one spec against a Root
the first run had closed, writ verify and readback never closed their
loading environments, devlore-test aliased the spec's root into
TestContext -- but they were not what those eighteen tests were failing
on. A natural experiment inside the plan package proved it: the three
helpers that close their environment had no failing tests; the five that
did not accounted for exactly fifteen.

A right number reached by wrong reasoning is still a miss, and the
campaign's counts are only as trustworthy as their attributions, so both
plans record it rather than restating the target. env-minted-root.md is
marked complete with the three-point measurement table; the 48-row in
platform-test-matrix.md now points at the correction instead of the
superseded diagnosis.

Remaining 28: file-provider write/permission semantics, path semantics,
chown, git argv, 2 CLI error-text expectations, and singles including
#376's Starlark path escape.

Assisted-by: Claude
Signed-off-by: David Noble <[email protected]>