Add a `test-author` skill — operationalize the test-contract gate for matrix YAML + server_behavior

[jonathaneoliver]

Motivation

The test-contract gate now lives in characterization-principles.md + an always-on memory (state intent, quiz on blanks, echo resolved config before any run). A skill turns that convention into a walked procedure that carries the authoritative knob vocabulary, so the intent quiz is concrete and the echo-back is a filled spec — not prose. The config surface (tests/characterization/matrix/*.yaml — 20+ arms, namespaced is.*/proxy.* knobs; plus server_behavior params) churns constantly, which is exactly when a wrong assumption silently builds the wrong test.

Scope

New .claude/skills/test-author/SKILL.md covering two families:

• char-matrix YAML (tests/characterization/matrix/*.yaml) — class config|fault; the full is.* (client, cold-relaunch) / proxy.* (server, config-on-connect) knob reference from internal/charmatrix/spec.go; axes: cartesian vs groups/compare A/B.
• server-behavior (tests/server_behavior/server_*_test.go) — the baseline-plus-tolerance contract shape + sb_common_test.go helpers.

Procedure: classify family/class → contract quiz (behavior · platform · constant target · what varies · pass/fail DATA signal) → draft → echo resolved config → --dry-run verify, before any run.

Boundaries

• Defers correctness rules to characterization-principles.md (links, doesn't restate).
• Follows CONVENTIONS.md (no-guessing, pass-signal-is-data, Bash-first).
• Out of scope: running tests (make characterize-* / harness char matrix), the unattended loop (sweep), analysis (forensics), and Roku.

Acceptance

• Skill is invoked on "write/add a characterization test / new matrix / server-behavior case / A/B comparison" and NOT on run/analyse asks.
• It produces a spec, echoes the resolved knobs, and ends with the exact --dry-run / -run verify command.

[jonathaneoliver]

Points: 3 (Fibonacci)
Effort: ~3–4 hours
Priority: P2 — internal tooling quality, no production impact; queue behind current sprint work
Value: medium — dev-velocity win (prevents silent wrong-assumption test builds) but not user-facing

Rationale: Single .claude/skills/test-author/SKILL.md file. Scope is fully bounded: lift knob vocabulary from internal/charmatrix/spec.go + sb_common_test.go, write the classify → quiz → echo → dry-run procedure, link (don't restate) characterization-principles.md and CONVENTIONS.md. No code paths, no cross-service wiring. Bulk of the work is reading the right source files and making judgment calls on what vocabulary to include — solidly half-day, not a feature. Sized against #838 (points:8, real cross-platform feature) as a reference: this is well below that bar.