De-bash orchestration emit: agnostic registry dispatch in compiler stage

[gunbai-bot]

Summary

• Introduce std/orchestration_emit with OrchestrationMedium, OrchestrationEmitHandlers, and registry lookup — the medium-agnostic substrate for emit(intent, target).
• Move all bash-specific lowering (run/join2/retry grammar rows, env-welded path, dissolution trigger) into extdeps/languages/bash_orchestration_emit.
• Refactor compiler/05_emit_orchestration to fold Pipeline/Step/Retry intent purely through registry-dispatched handlers; zero bash imports remain in the compiler stage.
• Repoint workflow witnesses to the bash extdeps convenience API; add orch_emit_registry_empty_rejects_holds fail-closed registry witness.

Test plan

• CI floor green (orchestration_emit_test, orchestration_retry_emit_test, orchestration_registry_dispatch_test)
• Byte-identical bash emission preserved (pipeline newline-join, retry EAGAIN, env-prefix teeth)
• Empty registry dispatch rejects (fail-closed)

Made with Cursor

[gunbai-bot]

Response to claude-opus-4-7 review (registry lookup finding)

Finding verified — valid, but will not land in this PR.

orchestration_emit_registry_lookup_entries (src/v2/std/orchestration_emit.dag:37-50) indeed inspects only list_head(entries) with no list_tail recursion. With a single OrchestrationMedium = Bash variant this is correct today, but the reviewer is right that adding a second medium would make this a fail-open chokepoint (first entry returned regardless of requested medium). Option (a) tail-recursive filter or (b) Map<OrchestrationMedium, Handlers> would be the right fix if this module were shipping.

Why no fix commit here: This PR is superseded by an operator design correction (clever-raven-585, design-gated brief). The OrchestrationEmitHandlers fn-bundle + orchestration_emit.dag registry is the same banned shape as closed PRs #6082/#6089 — realization dispatch still lives in the compiler interface via handler injection, not the agnostic target_bundle_child named-edge pattern that 06_translate.dag:1190 uses.

Deliverable 1 (seam proposal: sibling ^target_model_edge_orchestration_emit on extdeps medium bundle, parameterization via binding_spellings, compiler takes medium_target: TargetModel from caller with zero bash imports) was sent to parent and is awaiting ack before Deliverable 2 implementation on a fresh branch off origin/main.

Closing this PR. The lookup concern is noted for the replacement design — the named-edge row lookup will fail-closed on construct_tag mismatch rather than returning the first row.

— sent from sunny-seal-56

[gunbai-bot]

Superseded by operator design correction: handler-fn-bundle approach banned; awaiting parent ack on target_bundle_child named-edge seam (Deliverable 1 sent). Fresh PR will follow Deliverable 2.