W5-A: plan-driven prompt rendering, bounded codebook search, one JSON serializer, nonce delimiter

[bnsd55]

Fixes GPT Pro round-2 findings 1, 2, 39, 44 (scope per W5-A: schema.py prompt rendering + codebook, engine _prefill delimiter). Branch off origin/main 1f9f453. PROMPT_VERSION v7 → v8.

Finding 1 — the prompt and the compiled codebook disagreed (P0)

to_schema_str(mode="slots", tokenizer=...) now renders the displayed aliases from the compiled slot plan (plan['fields'][name]['aliases']). The independent _alias_code(i) prompt path is deleted — the tokenizer-free slot rendering raises instead of drifting. When the search picks digits, the prompt now shows 0)", 1)"..., so the model is taught the protocol the scorer judges. to_alias_schema_str(tokenizer) takes the tokenizer (engine _prefill, openai_slots — threaded through decide_openai/_decide_scalar_field/_decide_multi_field/_user_content, cli). alias_for_index survives only for the OpenAI adapter's per-field rows (documented at the definition).

Finding 2 — greedy codebook search rejected satisfiable fields (P0)

_greedy_pick → bounded backtracking over the prefix-conflict graph of pre-tokenized candidate rows. The failing tokenization (A a token-prefix of B/C/D, mutually prefix-free B/C/D, n=3) no longer raises: the search finds {B, C, D}. Every complete set is scored with the existing lexicographic objective (trie nodes → max depth → max tokens → length variance → code length). Bounds: 20k visited nodes per pool (2k for n>16), remaining-feasible prune, and an early exit when a set reaches the objective's first-component minimum (1 branch node) — keeps the 26-choice compile <1s. Deterministic (fixed pool priority order).

Finding 39 — one canonical JSON serializer

New jevmlx/json_text.py (json_text, ensure_ascii=False) used by every prompt and candidate path: schema.py labels candidates previously used stdlib json.dumps (escaped \u00e9) while the prompt showed the real character — scorer judged tokenizations the model never saw. All prompt/candidate call sites in schema.py, engine.py, openai_slots.py converted.

Finding 44 — ambiguous context delimiter

_context_block(context) / _context_nonce(context): both fences carry a sha256-derived 16-hex nonce of the context (<<<CONTEXT:C<nonce> … CONTEXT:C<nonce>>>), so a context containing CONTEXT>>> can no longer close the block early. Used by the naive path, the parallel prefill, and the OpenAI adapter. Formatting correctness, not a security boundary.

Verification

• tests/test_w5a.py: 10 tests, all FAIL on 1f9f453 (checked by checking out base), all PASS here — covering each finding's exact failing case.
• Full suite: 558 passed (not-slow), ruff check + format clean.
• Perf guard kept: 26-choice codebook compiles <1s (existing test updated only where behavior was the bug's).

Notes

• Findings verified in code before fixing; none of the four was wrong.
• No fallbacks/dual paths: the old rendering and delimiter are gone, not gated.
• Out of scope here (W5-B/C per the wave plan): findings 3–13, 18–28, set solver, decide_many parity.