docs: state the cap as three counts, because one number hid a disagreement

[hunterdsp]

Taking the plan README correction, per the review thread on just-buildit/just-makeit#636. The matching fix to the #555 body is already done.

What was wrong

Cap: 38 targets with every flag on was not wrong so much as unfalsifiable — and it had already failed silently. The P0 prototype also measured 38, and the two were different sets of 38:

	plan	P0
lint-<tool> dispatch	counted (~3)	excluded
standard-check, help-check, ghost-check	not counted	counted (3)

Two offsetting errors landing on one number is how a cap stops being a check. It agreed by coincidence, so it would have kept agreeing while both sides drifted.

What it says now

35 user-facing + N dispatch + 3 enforcement, as three counts, because they scale differently: N is per-repo (the tools that repo configures), the enforcement gates are fixed, and only the 35 is a property of the standard itself. Repo-local targets are outside the cap by definition — they are what criterion 1 bounds.

The 35 is recorded as the #555 group table summed (9 universal + 2 aggregates + 24 across the nine groups), with the per-group breakdown inline, so a build can be checked against it target-for-target rather than in total. That matters: summing to the right number is precisely what hid this, and checking the table row-by-row is what caught HAS_COVERAGE missing coverage-gate and HAS_EXAMPLES carrying an examples-clean that was never in it.

Verification

The repo's own hooks (mdformat 1.0.0 + gfm/gfm-alerts/mkdocs, as pinned here) run clean over README.md with zero rewrites. profile/README.md is untouched, so check-readme.yml is unaffected. No phase checkbox is ticked here — P0 is green but not yet merged; that tick belongs with the merge.

[hunterdsp]

Ticked P0 — just-buildit/just-makeit#636 merged as 7cb1e8a, and main verified green after it (gates pass, 33 targets documented).

Folded into this PR rather than a third one, since it is the same section. Also recorded the two faults the prototype caught before publication — the GNU make 3.81 macOS ships, and a make -p database whose wording changed in 3.82 — because both were in the file every adopter would have vendored verbatim, which is the argument for this phase order over publishing first.

P1 publishes the post-review file, not the one P0 opened with: it also carries coverage-gate, HAS_EXAMPLES reduced to test-examples, and all defaulting to build where HAS_C.