docs: rewrite README and npm page around what the gate catches

[EivMeyer]

Summary

Rewrites README.md and the npm page at packages/opcore/README.md. No code changes.

Both pages opened with "changed-file validation gate", which is the category husky, lint-staged and every pre-commit config already occupy, and then spent the hero example on a TS2322 that tsc --noEmit already reports. The Rust graph, which is the only claim on the page a competitor cannot copy, was a subtitle clause beginning "Backed by".

What changed

Lead with the gate, not one check family. The hero is now the whole pass and the single exit code. An earlier draft of this led with dead exports instead, which was worse: TS_DEAD_CODE_UNUSED_EXPORT is warning severity and exits 0, so headlining it would have promised blocking behaviour the check does not have.

The example shows the mix. One blocking FAIL next to two graph-only WARNs, severities visible. The type error demonstrates the gate is sufficient; the dead export and the missing TESTED_BY edge demonstrate what a per-file linter cannot reach.

Introduced report mode is documented. Neither page mentioned it. check --changed defaults to --report-mode introduced (packages/opcore/src/check.ts:167, and the CLI help says so at :246), so a pre-existing error in a file you touched does not block you. That is the reason the gate survives contact with a repo that has debt, and it was invisible.

Checks grouped by decay type, not by language. Structure, untested surface, duplication, size and complexity, documentation, correctness. The old table sorted by stack with a Deep / Useful / Experimental column, so a Python reader met "Experimental" before meeting any value. Stack maturity survives as a small reference table with counts instead of adjectives.

The documentation checks exist now. Ten of them, named, labelled opt-in. They were absent from both pages. clone.duplication was too.

Numbers. "41 checks, 25 in the default gate." Neither page carried a quantity.

Known limits, five distinct failure modes. Graph must be built first, docs checks are opt-in, higher-order usage reads as dead (#299), public entrypoints need declaring, Python maturity. A specific admitted flaw with an issue number buys more trust than the four denial sentences it replaces.

Removed. "not for remote publishing or opaque ratings", "does not publish, install packages, run wrapper tools, or edit source files", "does not invent findings or ratings", and the "Providers assess; ASP hosts decide" slogan. Four denials of accusations nobody made. Also "Metric output is named evidence and deltas", which parses to nothing, and the term "evidence" everywhere it was not naming a graph edge.

Added opcore graph build to the command list. On a fresh repo I got zero structural findings until I ran it, then nine. Real first-run cliff, undocumented.

On the count

Counted against the check-ids.ts arrays on main @ ab887a7: typescript 10, rust 11, python 9, docs 10, clone 1 = 41.

Worth flagging: python grew from 6 to 9 when python.ruff-lint, python.ruff-format and python.pytest arrived with #258 and #259. Anything still saying 38 is counting against af09b7c.

The default gate is 25, taken from an actual check --changed manifest rather than inferred: { typescript: 7, rust: 11, python: 6, clone: 1 }. The 16 opt-in checks reconcile exactly: 10 docs, typescript.lint, two config-gated TypeScript checks, and the three new Python ones, all carrying defaultScopes: [].

Verification

• tests/launch-claim-scrub.test.mjs 8/8
• npm run release:hygiene passes
• tests/package-identity.test.mjs 5/5. This one caught a real mistake: my first draft reworded the ASP provider section and dropped the required "opcore package exposes both" phrasing, which the test asserts against both this README and docs/quickstart.md. Restored.
• Every check ID in the new copy grepped against packages/validation-*/src/.

Open question

I regrouped into six buckets rather than the deck's CLEAN / CONTEXT / SHAPE, because three buckets made "Shape" and "Clean" carry too much. If the deck and README should use identical language, say so and I will collapse them back.