docs: anchor the shell → intent rescope (§3/§4) in DESIGN.md

[gunbai-bot]

Anchors the shell→dag rescope the operator and I aligned on 2026-07-17, per your request to put it in DESIGN.md with real .dag references.

What this says

The .dag intent may not name a target language. A workflow is an ordinary dependency graph over modeled operations (§4: a program is Node+Edge) — the intent is the .dag graph, not a Pipeline (that's optional sugar). Rendering to bash is a separate, target-parametrized concern, and it splits into two distinct layers that we'd been conflating:

• emit (§4): emit(intent, Bash) — one grammar read backward, target as a parameter.
• realization (§2): pure-spec → host-effect, N transports (LocalShell/SshShell) — consumes emit's output.

Bash ends up exactly where Rust is: its grammar spec (src/v2/extdeps/languages/bash.dag) and the emit rows, never in the intent.

The enforceable invariant

The intent imports no language-construction vocab (bash_build, the ShellStmt coproduct, serialize_bash) — that's emit-internal. This is realization_vocabulary_containment generalized with the intent layer in scope; green = conforms to §3/§4. Not a vibe — an executable wall.

Why now

Names the residual flagged 2026-07-17: (1) raw concat shell (gunbc_ci_deploy_invoke — untouched by the bash_build migration, which only moved the ShellStmt/program.dag consumers) and (2) structured-but-still-bash (floor_diff_observe). The sidecar delete (PR-A/PR-B merged) is demoted to Phase 0 of a bigger arc; git.diff/floor_diff_observe is the flagship for intent→emit→realization.

Contents

• DESIGN.md: one Open-threads anchor bullet (real .dag references throughout).
• docs/plans/shell-intent-emit-realization-design.md: the full design — layering table, invariant, residual, rescoped phases, coupled HostEffect.ShellCommand{script} lane, non-goals.

Docs-only. A residual-count census is running; exact per-class/per-layer numbers will land as a follow-up commit into the plan doc's census section.

🤖 Generated with Claude Code

[gunbai-bot]

Both findings addressed in a42b3a943a — thanks for catching them, especially the anchor one.

Finding 1 (DESIGN.md anchor reverted / git diff empty): Root cause is that DESIGN.md is a generated artifact — dag/gunbc/design_document.dag expected_design_md() (the DesignArtifact in the generated-artifact registry) renders it, and the drift gate enforces committed DESIGN.md == expected_design_md(). My first pass hand-edited the generated output, which correctly drifted; the WIP commit you saw was an auto-commit of my revert while I moved the content to the real source. The anchor now lives in design_document.dag's open_threads_blocks() as an li(...), and DESIGN.md is its rendered form — both carry the bullet, so the diff is no longer empty and the drift gate is satisfied (CI re-running to confirm byte-match).

Finding 2 (docs/plans/...:44 undefined Class 3): Fixed — relabeled Class 1 + 3 to Class 1. The census's fragment-helper sub-kind (flag/path string builders) is a raw-String builder, i.e. Class 1; there was no separate Class 3. The residual section now reads consistently against the ~110 figure.

— sent from calm-ferret-849