W5b-1: immutable schema (C7) + no compat re-exports (C13)

[bnsd55]

Implements GPT-REVIEW-2 section C items 7 and 13 on w5b-immutable, rebased on origin/main 649ba70 (post-#37/#38/#39 merges).

C7 — StructuredSchema / FieldDefinition immutable after construction

• FieldDefinition is a frozen slots dataclass: every attribute set exactly once at construction; mutation raises FrozenInstanceError (dataclasses.FrozenInstanceError via the slots descriptors).
• Deep-frozen at the boundary: choices is a tuple (copied from the caller's list), choice_descriptions a MappingProxyType over a copy, set_constraints a tuple of MappingProxyType constraint views — validated copies, nothing caller-mutable reachable.
• compile_set_constraints(constraints) validates exactly as before (same SchemaCompileError set: unknown types, non-multi fields, implies cycles, cross-group contradictions, exact_k=0 exclusions) and returns a NEW frozen field (validated deep-frozen constraints); StructuredSchema.__init__ swaps it into the dict. compile_set_constraints(None) returns the same field. compile_set_constraints itself is otherwise untouched for coder3's W5-C.
• StructuredSchema.fields is exposed as a read-only mapping (insert/delete/clear/pop raise TypeError/AttributeError); __getitem__/__len__ unchanged.
• Compiled plans (compile_slot_plan/compile_labels_plan and their cache) are deep-frozen once at compile time (_freeze_plan: dicts→MappingProxyType, lists→tuples). Cache identity is preserved — the cached object IS the frozen one, so cache-hit tests still assert is. Consumers see identical content; the engine's row-assembly concatenations now wrap plan reads in list(...) (4 sites).
• plan_hash now thaws the frozen plan (mappingproxy→dict) before hashing: the stdlib default=list path silently collapsed mappingproxies to their key lists, which hashed slots and labels plans identically — the existing mode-sensitivity test caught this.
• to_dict() unchanged in shape (returns plain copies).

C13 — api.py: zero backward-compatibility re-exports

• Deleted the compat re-export block and __all__ entries for resolve_model/MODEL_ALIASES (jevmlx.models is the import home; ARCHITECTURE rule "no backward compatibility" stays). api.__all__ now lists exactly the API's own surface (decide, decide_many, schema_from_model, Decision, FieldResult, NONE_OF_ABOVE, NONE_OF_ABOVE_DESCRIPTION, DEFAULT_MODEL).
• Dropped api.py's unused MODEL_ALIASES/resolve_model imports; updated the one stale docstring pointer (jevmlx.api.resolve_model → jevmlx.models.resolve_model in engine.py). No callers referenced the deleted re-exports (verified repo-wide).

Tests

• New tests/test_w5b1_immutable.py (12 tests): attribute/choices/descriptions/fields mutation raises; caller dicts not shared; set-constraints frozen + caller copies; compile returns new frozen field / same field on None; plans read-only incl. count sub-plan; cache identity on frozen plans; C13 __all__ contract.
• Updated existing plan-shape assertions to the frozen (tuple) contract (test_trie.py, test_multi.py, test_w2e_rowcodes.py, test_w2c_codebook.py, test_w5a.py, test_api.py, test_w2_setcons.py).
• 596 passed (not-slow), ruff check + format clean. Slow suite: the one failure (parity_real_model_twin, 0.071 vs 0.05s perf bound) reproduces on clean origin/main — pre-existing, not from this branch.

Coordination

• compile_set_constraints touched only to freeze its output (per the W5-C/coder3 note); its validation logic is byte-identical.
• Rebased on origin/main 649ba70 after W5-D's engine changes; the single conflict was my docstring pointer against the rewritten _load_engine_resolved docstring — resolved by keeping W5-D's text and re-adding the pointer.