Refactor dialect front ends behind a PlxSurface.parse_body vtable

[jonahharris]

What & why

The transpiler kept every dialect's lexer, parser, and emitter inside the single ~11k-line src/plx_transpile.c and dispatched them with a hardcoded block_style if-ladder — the "general" transpiler owned each language's implementation. This restructures it to mirror the precomp design: a dialect-neutral engine plus per-dialect front ends selected through a function pointer on the per-dialect descriptor.

This is a pure architectural change. The generated plpgsql is byte-identical.

Architecture

• PlxSurface becomes the vtable. It already rides on cx->surf; it gains void (*parse_body)(struct PlxCtx *cx) and a self_contained_block flag. plx_transpile() now calls cx.surf->parse_body(&cx) and runs a single assemble tail (conditional on self_contained_block for PL/SQL, which emits its own DECLARE/BEGIN/END). The block_style if-ladder is gone; block_style survives only as a lexer hint (indent vs. newline tokenizing).
• New src/plx_engine.h exposes the shared engine: the context (struct PlxCtx), Tok/PlxLocal2 types, PLX_MAX_DEPTH/PLX_DIAG_*, and the plx_*-prefixed entry points (plx_lex, plx_rewrite_expr, plx_emit_core/leaf, plx_span_text, plx_local_add/find, plx_skip_seps, plx_is_ident[_start], …). The engine — lexer, expression rewriter, leaf emitter, symbol table, interpolation, assemble — stays in plx_transpile.c, now ~2.6k lines.
• Each dialect owns its front end. COBOL/PL-SQL/T-SQL/Go/Ruby/Python tokenizers+parsers move into their plx_dialect_*.c; the shared PHP/JS/TS brace parser (plus TypeScript preprocessing) moves into a new src/plx_parse_brace.c. php/js/ts become thin config stubs whose parse_body points at plx_brace_parse_body.

Result

	before	after
plx_transpile.c	11,135	2,655
per-dialect files	~70-line stubs	478–2,277 lines each
new files	—	plx_engine.h, plx_parse_brace.c

Adding a dialect is now: new plx_dialect_X.c with a PlxSurface + parse_body, register it — no edit to the shared engine.

Verification

• Byte-identical output: md5(pg_proc.prosrc) captured for all 209 functions the regression suite creates, and diffed after every relocation step — identical throughout.
• All 13 installcheck tests pass at each step and at the end.
• Zero compiler warnings on a full clean rebuild.
• No transpiler logic was changed — only relocated.

🤖 Generated with Claude Code

[jdatcmd]

Approved — verified locally

Reviewed as a pure architectural refactor and verified the claims empirically (build + full regression suite against unchanged goldens), not just by reading the diff.

Verification

• Clean rebuild of all 14 translation units: zero warnings.
• installcheck (13 tests, all dialects): 13/13 pass against the unchanged test/expected goldens — consistent with the "no functional change / byte-identical output" claim.
• The one genuinely new logic surface — the plx_transpile() driver — checks out as a 1:1 behavioral equivalent of the old dispatch: the new self_contained_block path reproduces the old PLX_BLK_PLSQL branch exactly, and the else path reproduces the old DECLARE/BEGIN/END assemble exactly. The block_style if-ladder is correctly reduced to a lexer hint.
• Diff touches only src/* + Makefile; no test/expected edits, as expected for a no-op refactor.

Assessment

• Mirrors the existing precomp engine+front-end design, so it's consistent with the established architecture.
• Real maintainability win: plx_transpile.c 11,135 → 2,655 lines; adding a dialect no longer touches shared code.
• Low risk: mechanical relocation with a clean, verified dispatch equivalence.

LGTM. 🤖 verified with Claude Code.