Measure whether /tdd:principles actually gets invoked during implementation (SkillUse telemetry, not evals)

[kyle-sexton]

Spun out of Lane C (#2936). Refs #2933.

The claim, and why it is still open

#2936 asserted that "/tdd:principles exists but rarely gets invoked during implementation (known behavioral gap)."

While adjudicating that lane I first judged the claim STALE, on the grounds that the invocation is already wired — implement/SKILL.md names it at the "Before writing first test" chaining row, and grep finds 7 wired sites across the fleet. Validation showed that inverts the claim. Wiring is a property of text; invocation is runtime behavior. Seven wired sites plus rare invocation are not a contradiction — jointly they are evidence that wiring is not the lever. The presence of the instruction is exactly what makes a behavioral gap interesting.

Corrected disposition, now recorded in docs/upstream/aihero-shipping-course.md (Lane C): UNVERIFIED, with wiring confirmed at 7 sites and the behavioral half unmeasured.

Why not evals

The first disposal proposed was "file it as an evals item." That venue does not execute anything. plugins/evals/README.md:24-26:

No command in this plugin executes model-graded evals.

CI schema-validates eval content only. Filing there would produce JSON no runner runs.

The instrument that does exist

plugins/claude-ops/hooks/skill-usage-audit.sh writes a SkillUse event per skill invocation to skill-usage.jsonl, with a sibling skill-usage-expansion-audit.sh covering the expansion path (source: "tool" vs source: "expansion").

Recorded fields today: ts, event, skill, branch, project, project_id, hook, source.

Stated limit, up front: there is no caller attribution in that record. Nothing says which skill invoked tdd:principles, or that an implementation session was underway. The best available reading is co-occurrence — implement and tdd:principles events on the same branch within some window — which is a proxy, not proof. Any conclusion drawn must carry that caveat, and inflating the proxy into a measurement is the specific failure mode to avoid here.

What to build

1. A query/report over skill-usage.jsonl that answers: across sessions where implementation:implement fired, in what fraction did tdd:principles also fire, and in what order. plugins/claude-ops/skills/observability is the natural home.
2. The answer recorded against Lane C: TDD wiring + zero-assembly default chain #2936's premise in the SSOT — either the gap is real (and then the remedy is a separate item, since more wiring is already ruled out as the lever) or it is not.
3. Optional and worth costing separately: adding caller context to the SkillUse record so co-occurrence can be replaced with attribution. That widens a telemetry schema and touches the hook-telemetry convention, so it is a decision, not a detail.

Acceptance criteria

• A repeatable reading of implement→tdd co-occurrence exists, with its proxy limitation stated in the output itself
• Lane C: TDD wiring + zero-assembly default chain #2936's premise is settled in the SSOT as measured, with the number, or explicitly recorded as unmeasurable with the current schema
• If unmeasurable, the schema-widening decision is filed as its own item rather than assumed