Phase 3: entities → ProblemSpec projection; amicode_formulate/solve v2 emit scriptless solvespecs

[aarontrowbridge]

Roadmap context: vault plan `plan-20260812-142100-problemspecs-issimo-roadmap` (W2.1); master design `spec-20260717-020639-typed-problem-specs-piccolo-amicode` ("Amicode integration" + the entities↔spec mapping table). This is the reliability payoff of the whole typed-specs initiative: zero LLM-authored Julia for the spec-expressible subset. Depends on harmoniqs/Piccolo.jl#271 (shipped env admitting v1.21).

What exists today

solvespec v4 already routes `problem_spec` → `Piccolo.Specs.solve_spec` via `local_executor.ts` (`PROBLEM_SPEC_RUNNER`), and the gate already skips the import scan for scriptless specs. What's missing is the authoring side: nothing in TS produces a ProblemSpec from interview state.

Tasks

• Deterministic projection (no LLM in the path): `packages/schema/src/project.ts` (new) mapping `FormulationEntity` + `CompositeSystem` → `ProblemSpec` per the master spec's mapping table (parameterization → pulse.kind + template; min_time → `goal_treatment="both"` + `free_dt` + time objective; ensemble robustness → sampling wrapper; leakage → options; `custom` objectives → "not spec-expressible" fallback with the reason surfaced).
• `amicode_formulate` v2 (opencode-plugin): on completion, project → `~/.amico/problems//problem.toml`, ajv-validated before write; entity TOMLs still written (UI unchanged).
• `amicode_solve` v2: reads `problem.toml`, applies overrides (`max_iter`/`T`/`N`), emits solvespec with `problem_spec` (path or inline); still never launches (the never-launches invariant holds; the AGENTS.md bash workflow owns launch for both tiers).
• Rename the workspace card `problem.toml` → `card.toml` (or teach `kindForFilename` a `[problem]`-table sniff): today the opencode plugin writes a card at that path while both validators map the basename to the `problemspec` kind — `amico-validate` on a workspace card misfires.
• Script-tier fallback: any facet combination the spec can't express reports the reason in the tool result and falls back to composed/free authoring (today's path, untouched).

Acceptance

• Master-spec success criterion 1: a CZ-on-transmons solve runs end-to-end from `amicode_formulate` + `amicode_solve` with zero LLM-authored Julia; final infidelity matches the script path within 1e-6.
• Criterion 2: an intentionally malformed spec is corrected by the agent in ≤1 retry using only the structured error payload.
• vitest goldens on the projection; ajv dual-validation extended to the emitted `problem.toml`.