Stable diagnostic codes, an ICU message catalog, and first error in source order

[luisleo526]

Summary

Lane CG-DIAGCODES, for the pineforge-app's 14-language diagnostics. Targets codegen 1.2.0: the public API grows, nothing changes in the emitted C++.

• Codes and args on every diagnostic. Every Diagnostic now carries a stable code (PF-E1203 / PF-W0412) and raw named args. That covers transpile_full(...)["diagnostics"], CompileError.diagnostics (support checker, analyzer, codegen, lexer/parser/limits, libraries, request.security contexts) and both transpile_json envelopes.
• How codes are assigned. No emitter call site changed. A lazy classifier (pineforge_codegen/diagnostic_codes.py) matches each diagnostic's English (message, hint) against the catalog's ICU templates. It runs on first access, after the transpile, so it never touches the C++. Literal text must match around the arguments, so a code is never guessed. A text no template renders gets PF-E0000 / PF-W0000, and the test suite refuses that.
• Catalog. pineforge_codegen/diagnostics_catalog.json (628 codes, about 240 KB, one code per line) ships in the wheel and the Pyodide archive. It is exported as diagnostics_catalog() / render_diagnostic() and as @pineforge/codegen-pyodide/diagnostics_catalog.json, and attached to each GitHub release.
  • Each code has severity, area, an English ICU message template and an ICU hint template, plus a one-line explanation.
  • Each arg has a kind: identifier | type | keyword | number | vocab (with its closed values) | text. text marks open English the transpiler builds, such as a nested reason. It's an addition to the requested kinds.
• Pin. tests/fixtures/diagnostic_codes_pin.json pins each code's meaning (severity, templates and arg names). Removing or repurposing a code fails CI, and adding one needs a catalog entry and a pin line.
  • scripts/gen_diagnostics_catalog.py extracts every template the source spells and --write adds codes. It reads emitter arguments, module tables, locals, callers' arguments, caught exceptions and helper return values.
• Code ranges. The first digit is the area: 0 source (lexer, parser, limits), 1 support checker and requests, 2 analysis, 3 request.security, 4 libraries, 5 codegen, 6 array/matrix history, 7 ta.*. Inside a switch arm the support checker reports its refusals as warnings, so PF-W1nnn (nnn < 500) is the switch-arm twin of PF-E1nnn.
• First error in source order (TOP's optional item). Since Emit checked strategy settings and contain C ABI exceptions #159, the settings metadata visited the input arguments (defval, options, minval, maxval, step) before the script body. An unknown name in a later input's minval was therefore raised ahead of an error on an earlier line. That error is now held until the body is generated, and the first error in source order is raised (tests/test_settings_error_order.py). No C++ changes.

Compatibility proof

• Byte identity. Main ff67081 vs this head, one fresh process per file, transpile_full first:
  • 325 corpus + 263 fixture sources;
  • C++ (plus inputs, strategyParams and requests) identical 550/550;
  • diagnostics identical 1218/1218 (level, phase, location, message, hint);
  • glue envelope entries identical 1218/1218, code/args aside.
• Coverage. No diagnostic in the corpus or fixtures falls back to PF-E0000 / PF-W0000. Across the pure-Python test suite's 669 distinct diagnostics, only 4 test-crafted texts (boom, …) are uncatalogued.
• Pre-existing nondeterminism. _security_expr_hist_series_names sorts (sec_id, id(node)), so the order of clear_security lines follows memory addresses. Main's own C++ for tests/fixtures/tail_a_tv/taila_nested_bucket_hull_lag.pine differs between two harnesses (6a2434… vs ea70de…). This PR keeps codes lazy so the transpile's allocations stay main's. It does not fix the bug (a follow-up).

Test plan

• tests/test_diagnostic_codes.py: catalog shape, pin, every spelled template coded, every corpus/fixture diagnostic renders back byte for byte, glue envelope.
• tests/test_settings_error_order.py.
• Rebased on main 463a9af (Emit immutable compiled execution capability receipts (capability extension v1) #167, capability receipts). Emit immutable compiled execution capability receipts (capability extension v1) #167 adds no diagnostic text; scripts/gen_diagnostics_catalog.py reports nothing uncoded.
• Remote suite on CLOUD against engine f3bf9f07adfa (the active baseline):
  • compile-corpus 314/314 (rj-20261004t121634-9ddff2);
  • Pyodide gate PARITY OK over 277 fixtures (ok=264/264, err=13);
  • pytest 5675 passed, 41 skipped, 1 failed, in rj-20261004t121634-9ddff2 and again in rj-20261004t124132-4e0b90.
  • The failure is tests/test_array_history.py::test_many_bindings_are_walked_once_each, a 40 s wall-clock budget. It took 53–55 s on the CLOUD spot VMs and 38.9 s in the previous passing CLOUD run (pre-Emit immutable compiled execution capability receipts (capability extension v1) #167 head). On the Mac, old main 3cc1cf5, new main 463a9af and this head time equal (7.06–7.18 s). A CLOUD control run of plain main 463a9af is rj-20261004t130450-5f347d.
• pr-gate PASS (no-improve-no-regression) for engine f3bf9f07adfa and codegen e81fe7ee2115, recorded (eventKey verdict-f3bf9f07adfa-e81fe7ee2115-80dc6517f2ac, receipt c6e96ed7…). Hard lane: 1009 probes, 0 regressions, rank sum 4030 → 4030. Target lane: 6980 probes, score 0, nothing changed.

🤖 Generated with Claude Code