Warn that Limited Data Set does not scrub an 835, and fix the suite counts

[IPerception]

The 835 gap

docs/835-validation.md already recorded that an 835 handed through Limited
Data Set comes back byte-identical — its patient identifiers sit directly
inside loop 2100 (NM1*QC, NM1*IL), outside the PERSON_LOOPS the rule
targets, so scrubLoop never reaches one. The doc called that "worth stating
loudly in the app's own UI copy: silence here would hand a user a file they
believe is scrubbed."

It was stated in the guide panel only. At the point of use the app said nothing:
load an 835, click Create Limited Data Set, download an unchanged file, and
nothing on screen distinguishes that from a successful scrub.

What changed:

• A caution in the Limited Data Set card (#deidRemit), shown by
  syncRunButton() whenever has835() is true of the loaded file. It names the
  reason rather than just the outcome.
• The post-run message names the reason too. An 835 previously fell through to
  the generic "No patient loops found — nothing to strip", which reads as
  already clean.
• The button stays enabled. Disabling it would read as "not yet", where the
  truth is "never, by design".

docs/835-validation.md now records that this is said, and where.

Doc drift

• README.md said eight browser suites; the runner reports nine.
• web/PERFORMANCE.md said four, and named an outdated subset — both predate
  the tree, claims, validate, deid and render suites landing.

Verification

Covered in render.mjs (section [5]) rather than the static lint, because the
failure mode is stale state — the caution not clearing when an 837 is loaded
over an 835 is the same bug pointed the other way. The test also asserts the
output is genuinely byte-identical to the input, so that claim is now pinned by
a test and not only by prose.

node web/tests/all.mjs     # 9 suites pass; render.mjs 26 checks, up from 19
python -m unittest         # 11 tests pass

Not yet exercised in a real browser — the DOM shim can't report how the caution
reads in the card or whether the 2.2s toast is long enough.

🤖 Generated with Claude Code