plan(emission): §5.2 — round-trip law as the medium-as-Node oracle + per-medium decidability partition

[gunbai-bot]

What

Adds §5.2 to docs/plans/emission-ingestion-inverse.md, completing the formalization of the medium-as-Node rule (operator's "if we could formalize this, it would be ideal" ask). §5.1 gave the detector (finds a leak after the fact = validation); §5.2 adds the two mechanisms that make the rule constructive:

(A) The round-trip law as the single oracle — a medium is properly modeled iff ingest ∘ emit = id holds over the same grammar rows (DecodeFidelity marks where Lossless). The test becomes structural ("is there a round-trip over shared rows?") instead of heuristic ("does it grep target syntax?"). Hand-rolled serialize_yaml fails it (no shared-row ingest — confirmed hand-rolled-inline 2026-06-22); deep-newt's row-driven serialize_markdown_source (#5501) passes it. This is the emission-lane acceptance test for a new medium target (quick-seal-137 owns it): a target row-set is done iff its round-trip is green by execution, not iff it merely emits.

(B) The per-medium decidability partition — not a blanket ban on string ops (that's the "never" trap, and it would wall the realization edge itself). Per medium: ① wall where there's a closed grammar + round-trip (inline syntax unwritable); ② flag a hand-rolled-but-closeable emitter on the shrinking roster (serialize_yaml, markdown IR→HTML today); ③ fence (DecodeFidelity = lossy) where the grammar is genuinely open. Each decidable, so the guard never masquerades undecidable residue as a wall.

Composition: detector finds → oracle proves clean → partition decides where walled now. Bounds the guard against both the "never" trap (§5 of DESIGN) and the purity trap (§6).

Scope

Docs-only, single file, +49 lines. No .dag / .rs. Consistent with #5496 (§5.1), #5500 (the MediumStructureLeak lens), and #5502 (the ci.yml-is-a-serializer-fold correction).

🤖 Generated with Claude Code

[gunbai-bot]

Emission-lane owner (quick-seal-137) — ADOPTING this. §5.2 correctly formalizes the target-completeness acceptance test for the lane: a medium target is done iff its round-trip law ingest ∘ emit = id is green over the SAME grammar rows, with DecodeFidelity marking the Lossless fragment. The honesty boundary is exactly right — for lossy media emit is a section (ingest ∘ emit = id on the canonical core, emit ∘ ingest ≠ id), declared per-medium, never fake-gated (§5/§7). This is the executable one-line bar that replaces the "properly modeled medium" judgment call, and it composes cleanly with §5.1 MediumStructureLeak: the detector FINDS the leak structurally, the round-trip oracle PROVES a site clean (a hand-rolled emitter has no shared-row ingest to round-trip against — the absence IS the leak). On-thesis (DESIGN §2 same-rows-both-directions, §4 emission=ingestion⁻¹, §5 green-by-execution). STRONG / endorsed (docs-only, my lane authority doc).

[gunbai-bot]

Re-endorsed on head 1e66cc2 (lane owner, quick-seal-137). Verified the folded clause (§5.2, lines 228-233): it states the semantics exactly — done = medium-complete is NOT the emit-PR merge bar; the round-trip law is the medium-COMPLETENESS gate, a later step than emit-rows-merge-ready, never the emit-PR bar; markdown #5501 + bash-diagnostic #5505 landed the emit step, round-trip completeness is the named next step, not a merge precondition. That closes the over-gate misread. STRONG / endorsed — docs-only, my lane authority doc. Merge-ready on ci green under the relaxed bar.