feat(skills): add the wtf-is-real verification skill

[Nuropunk]

What this adds

A new public, installer-facing skill at skills/wtf-is-real/, built around one rule: never accept a claim's self-report, find the signal that could not exist unless the claim were true.

The teaching example is the sharpest form of the idea. A timer sitting at zero refutes a claim that a recording is running, but only if the timer is fed by the encoder's sample count. A timer fed by the wall clock counts up forever next to a dead recorder, so it manufactures confidence instead of evidence. Same widget, opposite evidential value. That gives the one question the skill applies to every check it is offered: if the work stopped right now, would this signal change?

Structure

SKILL.md owns the operating contract:

• Two gates. The REAL gate (did it happen) runs before the MATTER gate (was it worth happening), and passing one is never permission to skip the other.
• A 0-6 evidence ladder from assertion through self-report, self-check, artifact, out-of-band observation, negative control, and independent re-derivation. A verdict carries the rung of its weakest load-bearing evidence, nothing ships on rung 2 or below alone, and anything irreversible needs rung 5.
• Depth scaled by stake, not curiosity, so a reversible typo does not get a courtroom.
• The kitchen audit's four counters: ingredients, recipes, supply chain, and the plate. Most verification checks the plate and never checks whether the flour was flour.
• The clipboard, a fixed verdict format carrying the claim, rung, discarded signals, negative control, what is absent, the remaining gap, and one actionable sentence.

Four references carry the depth:

File	Contents
references/verification-methods.md	Twenty-two methods grouped by liveness, falsifiability, freshness, independence, causality, quantity, and provenance. Each states what it proves, how to run it, and its blind spot. Plus a claim-shape chooser and the verification-theater anti-patterns.
references/deputy-roster.md	Ten single-lens supervisors, the independence rule that stops a quorum collapsing into one opinion, quorum thresholds by stake, and the rule that silence is never a pass.
references/kitchen-audit.md	The four counters, per-counter procedure, and the audit sheet, where NOT CHECKED is an honest entry and writing PASS over an unwalked counter is the thing the sheet prevents.
references/pom-meaning-gate.md	The proof-of-meaning read. Defers to a canonical scorer when the host provides one, and never presents the fallback rubric as a canonical score.

assets/cheriff-mask.svg is the accompanying mask, self-contained with no external references.

Repo wiring

• All five prose surfaces registered in docs/documentation-audiences.json as public-product.
• The README's two-tier skill layout section now names the skill alongside skills/stow.
• No firstmate paths, tools, or vocabulary in any of it, per the skills/ tier contract.

Verification

Check	Result
bin/fm-doc-audience-check.sh	Passes: surfaces=74 local_links=267. Confirmed load-bearing by observing it fail first on the unregistered surfaces, then pass after registration.
Frontmatter	Parses as YAML with name, description, user-invocable.
Cross-references	All four intra-skill links and the one heading anchor resolve, checked by the audience check's link pass.
Mask asset	Parses as well-formed XML, contains no external references, and was rendered and inspected at 512px and at 48/72px avatar sizes.
One sentence per line, no em dashes	Swept across all five Markdown files.
bin/fm-lint.sh	Not applicable and not claimed: the diff contains no shell or workflow files, and ShellCheck is absent from this environment so the run was a no-op.

CI status: absent, not green

No CI ran on this pull request, and none will under the repository's current configuration.

Verified through the Actions API rather than inferred from the empty checks list: list_workflows returns total_count: 0, list_workflow_runs returns total_count: 0 across the whole repository, and the PR's combined status is pending with total_count: 0. Actions is not enabled on this fork, so the Require no-mistakes gate defined in .github/workflows/no-mistakes-required.yml, the behaviour suite, the lint job, and the repo-invariants job all simply did not execute.

An empty check list is not a passing check list. Treat this branch as unvalidated by CI.

Separately, the no-mistakes gate could not be run locally either: it is not installed in this environment and needs a local git proxy plus credentials. The deterministic signature it writes into a PR body is deliberately not reproduced here, since forging it is precisely the verification theater this skill exists to refuse. Landing this against a repository where the gate is enforced needs a git push no-mistakes from a machine that has it configured.