0.1d — Julia round-trip: validate the emitted result.toml + run-dir against the shared schemas (closes #18)

[Rchari1]

Stacked on #30 (0.1c). Do not merge. Completes the Phase 0′ SchemaPackage set.

Closes the producer/consumer loop: a Julia validator that checks result.toml/manifest.toml against the same packages/schema/schemas/*.json the TS side uses — read directly, no transcribed copy — so mutating a schema flips both languages. That's the cross-language anti-drift guarantee (#18 AC6), verified empirically in review (mutate a shared schema → both TS and Julia flip the same fixture).

What's here

• packages/schema/julia/validate.jl: AmicoValidate module + CLI over JSONSchema.jl (draft-07, the ajv↔JSONSchema.jl common floor). Reads ../schemas directly; parses TOML; field-precise errors via a shim that reaches TS parity — names the missing/extra key (missing required key "run_id", unknown key "rogue_key"), lists enum values, renders type/range prose (must be number, must be <= 1.0001), and the version-specific schema_version message. jsonify() coerces unquoted TOML datetimes (S2). Exit 0/64.
• packages/schema/julia/runtests.jl (22): valid golden + invalid corpus (field-precise) + schema_version policy + unquoted datetime + the real bundled demo run dir + the single-source anti-drift asserts (SCHEMA_DIR == ../schemas, no schema copy in the Julia tree).
• ci.yml schema-roundtrip job: setup-julia + instantiate + runtests.jl. Fast tier — JSONSchema only, no Piccolo solve (per the design-review tier split). Committed Manifest.toml for reproducibility.

Tier split (S1)

Fast = golden + negative round-trip (this job). Freshly-emitted-via-solve = slow/nightly extension; solve_template.jl + solve_common.jl already stamp schema_version, so emitted artifacts are round-trippable there.

Review

Adversarial review empirically verified the AC6 keystone (schema mutation flips both validators) and caught a must-fix (committed Manifest lacked Test → would red the new CI job) — fixed (re-resolved; fresh instantiate exits 0). Field-precise parity (#2/#3) + demo-golden (#4) folded in.

All four slices green: TS schema 32 / amico-run 47 / extension 63 + Julia 22.

🤖 Generated with Claude Code

[jack-champagne]

The schema craft here is solid and the parts that are present check out — Julia reads the same packages/schema/schemas/*.json the TS side does (real "no second copy" guard), and the schema-roundtrip CI job genuinely runs Julia and reds the build (not a skipped/false-green gate). I ran the suite locally: 32 / 47 / 63 / 22 Julia, all green.

Two changes before this closes #18 — both are "validate against the right thing," not "build more."

1. The round-trip validates fixtures, not the producer — which is the exact gap this slice exists to close. It validates committed test/fixtures/* + the committed demo/run/ dir. Nothing here runs a solve to emit a fresh result.toml and validates that. AC5's "or freshly-emitted ones" is unmet, and the "slow/nightly" lane the description points to doesn't exist (ci.yml is the only workflow). This is the β shape: an integration test that doesn't exercise the integration — it goes green while the real emitter is free to drift. Gate #18 on one real lane: a single canonical solve → run-dir → validate-with-validate.jl in CI. One prompt is enough. Until that exists, #18 stays open and the round-trip is validating its own fixtures.

(Architectural note: the orchestrator, not Julia, writes manifest.toml — so a Julia emit→validate can only cover result.toml. The title's "emitted … manifest.toml" overclaims; scope the producer round-trip to result.toml.)

2. AC6 "mutating the shared schema flips BOTH validators" is asserted, not exercised — and it isn't quite true. I tested it: added a required key to result.schema.json, then validated the clean fixture.

• Julia (runtests.jl) flipped — 3 tests failed. It reads the schema at runtime. ✓
• The shipped TS bundle did not flip (still OK). TS bakes schemas into dist/ at build time (import … with { type: "json" }); it only flipped after a rebuild.

runtests.jl's AC6 testset checks file-path identity + the Julia flip; it never exercises the TS side. So in any "edit schema, don't rebuild" window — including a shipped VSIX whose schema is frozen at package time — Julia and TS can disagree and this gate won't catch it. Either add a step that rebuilds TS then asserts both flip on the same perturbed schema, or document the build-time-vs-runtime asymmetry in the contract.

On scope — don't overbuild ahead of this. The fix above is one real lane, not more fixtures or more schema surface. Same for the stack: hold the env-contract (requires/provides) and any further validation machinery until the producer seam is actually validated. Validating the right thing once beats validating the wrong thing thoroughly.

Stack coordination: this hardcodes manifest.toml (validate.jl, fixtures, demo) — it must follow #28's manifest.toml → run.toml rename in lockstep or the round-trip validates a dead filename. And per #15's note, land #28 to main and rebase these off main rather than running a 4-deep chain.

Net: the validator is good; it's just pointed at fixtures instead of the producer. Repoint it at one real emitted run and this becomes the integration gate β never had.

[Rchari1]

Pushed addressing both must-changes:

1. Real producer lane (not fixtures). scripts/producer_roundtrip.sh runs amico-run on a minimal non-Piccolo emitter → a live run-dir (orchestrator writes run.toml+FINISHED, the Julia emitter writes result.toml) → validates the freshly-emitted artifacts with validate.jl. Wired into the schema-roundtrip CI job (fast tier; the real Piccolo solve stays the slow/nightly extension). Title scoped to result.toml per your note — the orchestrator, not Julia, writes the header.

2. AC6 exercised, not asserted. scripts/ac6_drift_check.sh perturbs the shared schema, rebuilds TS (the build-time-bake gap you found), then asserts both the TS bin and validate.jl flip on a previously-valid fixture, then reverts. Verified locally (AC6 PASS — both flipped); runs in CI.

Plus the manifest.toml→run.toml rename lockstep (validate.jl/runtests.jl). schema 34 / amico-run 47 / extension 64(+1 skip) / julia 22 green. Re-requesting.

[jack-champagne]

Re-reviewed 2ef1d88c. Both must-changes resolved — the validator now points at the producer, not its own fixtures:

• Real producer lane — producer_roundtrip.sh runs amico-run on a minimal emitter → a live run-dir → validate.jl on the freshly-emitted result.toml/run.toml/FINISHED. set -euo pipefail + validate exit codes make it red on drift; wired as a gating step in schema-roundtrip. Title scoped to result.toml per the note.
• AC6 exercised, not asserted — ac6_drift_check.sh perturbs the shared run.schema.json, rebuilds TS (the build-time-bake gap I flagged), asserts BOTH the TS bin and validate.jl flip on a previously-valid fixture, then reverts via trap cleanup EXIT. Closes the runtime-vs-build-time asymmetry.

Approving. Retargeting to main + merging behind #30.