live-matrix

The parameterized harness conformance scenario suite — eight live behaviors exercised by a test file against any registered launcher, filing per-scenario eval readings on that harness's node.

harness-adapter's acceptance rule — an adapter merges only with per-behavior readings measured through a REAL dispatched session — used to live as prose: each harness's eval.md hand-transcribed its own wording of the eight behaviors and a worker ran them by hand, so every new harness re-copied the matrix or silently dropped rows. This node keeps that coverage as a parameterized test asset: the behavior contracts live in each harness's eval.md scenarios, while one test file supplies the shared drive and evidence collection. Running the test against a launcher is a test action, not a new SpexCode CLI verb.

Each scenario carries its contract in eval.md and points at the test case that supplies the DRIVE (real steps over the public session verbs — new/send/show/stop/resume/close, plus a materialize for the transient guard hook and tmux for the liveness kill; never a parallel mechanism). The test captures EVIDENCE as a per-scenario transcript of every command, board observation, and pane capture, then files it with the reading. spec-cli/scenarios/harness-live-matrix.ts resolves the launcher to its harness, targets the <harness>-harness spec node (--node overrides), and walks ONE real worker through the whole lifecycle: undeclared-stop · pretooluse-block · ask-note · deliver-steer · resume · liveness · commit-gate · close-residue.

The scenario declaration is the contract source: the test resolves the existing canonical name or historical alias and fails loudly when the harness node has not declared it. It never creates or rewrites eval.md. Adding coverage means adding scenario data and, where needed, a parameterized test case; it does not widen the CLI or add a harness-specific route.

A harness whose declared runtime semantics intentionally remove a matrix premise does not fake the row. claude-headless is record-backed and has ephemeral turn children, so stop -> offline -> resume and SIGKILL -> offline are categorically the wrong measurements; its own idle-resume and record-liveness scenarios replace those two rows while the remaining shared behaviors and its interrupt addition stay live behavior readings.

Verdicts stay honest three ways: a row that could not be provoked (the worker declared on its own; no mid-turn window opened) files NOTHING and reports skip — never a fabricated loss signal; a measured row files pass or fail immediately, so an aborted run keeps its partial readings; and the runner's own board polling is exactly the probe pressure the delivery path must survive, so the measurement environment is the adversarial one. A new harness is covered by registering its launcher and adding its scenario data — zero new CLI code.