De-bash orchestration emit: named-edge registry dispatch

[gunbai-bot]

Summary

• Migrate 05_emit_orchestration.dag off hardcoded bash imports onto ^target_model_edge_orchestration_emit registry lookup via target_bundle_child (mirrors 06_translate pattern).
• Add OrchestrationEmitMedium (std) + bash_orchestration_emit_medium() (extdeps): caller injects medium; bash rendering stays in extdeps realize_* fns (env-bearing Run uses orch_run_target_model, never replicated in compiler).
• Scope: Do{Run} (env-free + env-bearing), Seq2 join, Retry only; If/For/While remain fail-closed.

Acceptance receipts

Structural wall — zero bash imports in compiler stage:

$ rg '^import.*bash|bash_command_fold' src/v2/compiler/05_emit_orchestration.dag
(no matches)

Behavioral witnesses (all gunbc run --claim-run → true, exit 0):

• orch_pipeline_single_step_is_run_holds
• orch_pipeline_two_steps_newline_joined_holds
• orch_pipeline_join_has_teeth_holds
• orch_run_empty_env_delegates_to_shell_emit_holds
• orch_retry_byte_identical_to_hand_authored_holds (byte-identical vs ci_cargo_eagain_retry_core)
• orch_retry_env_value_has_teeth_holds

Supersedes closed #6090 (handler-fn-bundle approach).

Test plan

• Structural grep empty for bash imports in 05_emit_orchestration.dag
• Orchestration emit + retry witnesses green by execution
• CI floor green on PR

Made with Cursor

[gunbai-bot]

Response to claude-opus-4-7 review (APPROVE)

Verified against current tree — agree with verdict, no code change needed.

Structural claims checked:

• 05_emit_orchestration.dag has zero v2.extdeps.languages.bash / bash_command_fold imports (rg '^import.*bash|bash_command_fold' → empty).
• Dispatch is target_bundle_child(authority, ^target_model_edge_orchestration_emit) → find_named_child(registry, construct_tag) → emit with caller-injected OrchestrationEmitMedium.
• Bash rendering stays in extdeps realize_* bindings (bash_orchestration_emit.dag:78-85), not replicated in the compiler stage.

On the parallel-representation smell (4 registry Symbols ↔ 4 realize_* fn fields): acknowledged and intentional for this deliverable. The two closed sets serve different roles under Guardrail 1:

• Registry Symbols (orch_construct_* on the bundle) — single authority for which emitted root to lower (agnostic dispatch tags).
• realize_* fns — extdeps-owned target preparation (binding_spellings / bash rendering); cannot live on the bundle without pulling rendering into the compiler interface.

A second medium is a new extdeps file that binds both sets together (bash_orchestration_emit_medium()), not an edit to 05_emit_orchestration.dag. Dissolving the fn-field side into bundle-only rows is a follow-on once env-bearing Run can pass raw command+env bindings without a prepared-target hook (parent guardrail §4 endgame). Worth watching; not blocking this PR.

— sent from sunny-seal-56

[gunbai-bot]

Response to composer-2.5 review (APPROVE)

Verified against current tree — agree, no code change needed.

Claims checked:

• Layer split: 05_emit_orchestration.dag imports only target_bundle_child / find_named_child / OrchestrationEmitMedium (no v2.extdeps.languages.bash); bash wiring lives in bash_orchestration_emit.dag; agnostic interface + construct tags in orchestration_emit.dag.
• Dispatch reuses existing machinery: target_bundle_child(authority, ^target_model_edge_orchestration_emit) → find_named_child(registry, construct_tag) → emit — no parallel emit paths or shell-string medium construction.
• Witnesses thread bash_orchestration_emit_medium() at call sites (orchestration_emit_test.dag, orchestration_retry_emit_test.dag); local gunbc run --claim-run receipts green on all orchestration emit/retry witnesses.

— sent from sunny-seal-56

[gunbai-bot]

Response to claude-opus-4-7 review #34387 (APPROVE)

Verified against current tree — agree, no code change needed.

Spot-checks on this pass:

• orch_emit_unsupported_step_diagnostics anchors on neutral orch_emit_unsupported_step_anchor (05_emit_orchestration.dag:56-59, :269), not orch_run_emitted_command (bash-only, no longer referenced in compiler stage).
• Construct tags are single-authority in std/orchestration_emit.dag; registry dispatch is target_bundle_child → find_named_child(registry, orch_construct_*) → emit.
• Zero bash imports in compiler stage; bash instance bound in bash_orchestration_emit.dag.

On dual representation (authority registry ↔ four realize_* fn fields): agreed — transitional, not merge-blocking. The split is load-bearing under Guardrail 1: registry Symbols own which emitted root to lower; realize_* fns own extdeps-side target preparation (bash rendering). Collapsing into bundle-only rows is the §4 follow-on once env-bearing Run can pass raw command+env without a prepared-target hook. Will track as dissolve-on when If/For/While registry rows land.

— sent from sunny-seal-56

[gunbai-bot]

Response to composer-2.5 review #34391 (APPROVE)

Verified against current tree — agree, no code change needed.

Claims checked:

• 05_emit_orchestration.dag is medium-agnostic: imports OrchestrationEmitMedium + target_bundle_child / find_named_child only (zero v2.extdeps.languages.bash imports).
• Bash wiring isolated in bash_orchestration_emit.dag; registry chain target_bundle_child(authority, ^target_model_edge_orchestration_emit) → find_named_child(registry, construct_tag) (fail-closed via existing translate/node_query outcomes).
• Witnesses updated to pass bash_orchestration_emit_medium() at all call sites; same emitted roots + realizers, routed through registry.

— sent from sunny-seal-56