0.1b — validate lab.toml on extension load with field-precise errors (closes #16)

[Rchari1]

Stacked on #28 (0.1a). Do not merge.

A partner lab's lab.toml is the only place hardware params enter a solve; β had no validation, so a malformed/mistyped config silently solved against the wrong hardware or failed opaquely mid-solve. This validates it on extension load against the shared @amicode/schema lab schema (defined in 0.1a) and surfaces a field-precise error (offending key + dotted path), non-fatally.

What's here

• src/lab_config.ts: resolveLabTomlPath (default ~/.amico/lab.toml, ~ expansion) + checkLabToml → absent | valid | invalid{errors}, via the single @amicode/schema validator (no second validation path).
• extension.ts activate(): validates on load; invalid → showErrorMessage with the first field-precise error + the full list in the "Amicode — runs" channel. Non-fatal — the rest of the extension still activates.
• amicode.labToml config (path; empty → ~/.amico/lab.toml).
• lab.toml.example: stamped schema_version = "1" so the shipped starter validates.

Tests (lab_config.test.ts, 13)

Field-precise negative matrix (missing / wrong-type / out-of-range / unknown-key / absent+unrecognized version), parity with @amicode/schema.validateFile (asserts the full errors array — proves single-path), and shipped-lab.toml.example conformance. Extension 63 green; typecheck clean.

Review

Adversarial code review: no must-fix / should-fix. Verified empirically that validateFile is throw-safe on missing/garbage/directory inputs (so the load seam can't break activation), and that lab.schema.json exactly matches the shipped lab.toml.example (a real install won't red on load).

🤖 Generated with Claude Code

[jack-champagne]

Code's correct — reproduced all 7 ACs against real ajv output; the load-time check delegates to the single @amicode/schema validator and is non-fatal. Approving the code.

Two nits that live in #28's schema (flagging here since this is the PR that owns the "no silent wrong-hardware" story): delta_GHz has no range bound while every other [transmon] field does — a sign-flipped or garbage anharmonicity passes validation, the exact class this slice exists to catch; and the only valid fixture is demo-lab, no Schuster profile (the PRD's demo forcing-function).

Tests + Phase 1: the negative matrix is partial vs #16 — it hits levels / drive_max / schema_version / unknown-key but not the range bounds (omega_GHz ≤ 100, drive_max_GHz ≤ 10) or lab.name minLength, and the "parity over the shared negative-fixture corpus" AC is exercised on a single input (inline). Two Phase-1 prerequisites this surfaces: (1) there is no test harness for the VS Code host layer — the on-load toast/channel path is untested, and Phase 1 stacks more UI (inspector multi-run, catalog tree) on that same seam, so an extension-host integration approach should land before it compounds; (2) lab.toml has no provides block yet — the cloud solve-service (aws-infra) keys on lab provides to reproduce the runtime and validate requires ⊆ provides, so that dimension + its validation is a cloud/Phase-1 prerequisite, not optional polish.

Sequencing, not a change: stacked on #28 (CHANGES_REQUESTED), and the tests hardcode #28's error strings — rebase + re-run after #28 settles, and don't merge ahead of it.

[jack-champagne]

[nit] this showErrorMessage + output-channel path is the one surface the unit tests can't reach — no VS Code host in CI, so checkLabToml is covered but the toast/channel wiring isn't. I'll verify it by hand on the packaged VSIX (bad ~/.amico/lab.toml → reload → toast + "Amicode — runs" list). Non-blocking — but it's the seam Phase 1's UI work compounds on, so it wants a host-level test approach soon.

[Rchari1]

Acknowledged — the toast/channel wiring stays unit-untestable (no VS Code host in CI). Noted as a Phase-1 prerequisite (a host-level/integration-test approach before the inspector/catalog UI compounds on this seam); not adding host-test infra in this slice. checkLabToml itself is covered.

[jack-champagne]

[test] this proves parity on one input; #16's AC says "over the shared negative-fixture corpus." Delegation makes drift structurally impossible so it's not urgent — but loop the corpus to meet the AC literally, and add the range-bound cases the matrix above skips (omega_GHz ≤ 100, drive_max_GHz ≤ 10, lab.name minLength).

[Rchari1]

Done — the parity test now loops a corpus (valid + missing/wrong-type/out-of-range/unknown-key/absent+unrecognized-version), asserting checkLabToml === validateFile on each, and the range-bound cases (omega_GHz≤100, drive_max_GHz≤10, the now-bounded delta_GHz, lab.name minLength) are added. Plus a Schuster-profile valid fixture (negative-δ, 4 levels).