2-emit-partition (b1a): v2 bash whole-program serializer foundation (direct-variant coverage)

[briansrls]

(b) bash grammar de-fork — foundation slice (b1a)

First slice of the bash de-fork (node 2-emit-partition (b)). The dsl fork extdeps.languages.bash.program hand-rolls serialize_bash(ShellProgram) -> String, consumed by ~17 modules; the v2 grammar authority (bash_command_fold) had no whole-program serializer (only per-construct builders). This adds one.

What this lands

• New authority module src/v2/extdeps/languages/bash_program_fold.dag: bash_program_serialize(ShellProgram) -> Outcome<String> — folds statements through the existing v2 grammar fold (bash_fold_stmt_emitted_witness + bash_fold_target_model_for_stmt_witness + emit), newline-joins with trailing \n per stmt, and applies the set -e prefix — matching dsl serialize_bash's shape exactly. Composes existing surfaces; no change to the fold mechanism.
• Byte-oracle test src/v2/test/claim/bash_program_fold_test.dag: bash_program_serialize == dsl serialize_bash byte-for-byte on a multi-statement program (set -e + command + assign), green by execution, plus a set_e-discriminating tooth.

Scope (honest)

Covers the direct target-model statement variants (Command/Assign/Exit/AndOr/Pipe/If/simple-Subshell). The delegated variants (RawLine/WithRedir/Heredoc/Negation/EnvPrefixed/Subshell-with-delegated-body) and the word gaps (Concat/CmdSubst/Raw, plus BraceGroup) currently Reject — their serialization machinery exists but is test-local and needs promotion into the authority. That + the ~17-consumer migration + breaking bash_command_fold's dsl self-import are the named follow-on slices (b1b / b2 / b3). This slice de-risks the architecture: proves v2 can serialize a whole bash program byte-identical to the fork.

🤖 Generated with Claude Code

[gunbai-bot]

All three findings addressed in the pushed fix — thank you, these were correct catches.

Layering (surfaced by CI, related): the module was in extdeps but imports v2.compiler.emit — an extdeps→compiler inversion. Moved to src/v2/workflow/bash_program_emit.dag (workflow may import compiler+extdeps; all imports now point toward std). Whole-program serialization inherently needs emit(), so its home cannot be extdeps; it's a language-specific workflow consumer, keeping compiler stages agnostic.

Finding 1 (medium-as-string): took the on-carrier scaffold-marker option you offered. Added bash_program_serialize_medium_join_scaffold naming the dissolution trigger — the set-e prefix and newline join become bash grammar productions emitted through the fold, mirroring the existing orchestration seq2 join (orch_construct_seq2 / orch_seq2_target_model), so serialization runs with zero concat. Tracked as the (b1b) follow-on.

Finding 2 (single authority): added bash_program_serialize_single_authority_reconciliation — bash_program_serialize is declared THE single v2 ShellProgram→String authority; serialize_bash is the scaffold-to-delete, dissolution = (b3) migration of its ~17 consumers + deletion. The coexisting pair is honestly-marked incremental de-fork debt with a named owner, not two co-equal authorities; the test oracle is the git-historical byte-reference the grammar surface is proven identical against.

Finding 3: facts_lookup is a fail-closed sentinel identical in shape to orch_emit_facts_lookup (05_emit_orchestration.dag:44) — direct-variant emit never consults facts (the oracle witness is green/Accepted, never Rejected). Renamed bash_program_fold_*→bash_program_emit_* for module consistency.

— sent from fierce-lark-783