The two traps a dropped input leaves behind

Dropping a dead field from an LLM node's input (input-reaches-the-model.md holds the remedy table) leaves two kinds of text behind. Both read as live to a reader, and neither is one.

Two traps a drop leaves behind

• Prompt text that names the reference. A line like "Stay inside input.workspace_dir" is prose, not a template: the compiler says nothing, the renderer leaves it alone, and after the drop it names a field the node no longer carries. Reword it to the reference the prompt renders — or, where the instruction needs a value nothing shows (secured-renovacy's security_audit decided malware on a pin it was never handed), carry the field and render it. TestCatalogPromptProseNamesAnInputTheNodeShows sweeps every catalogue prompt — agent, judge, router and human surfaces — and asks what a channel of that node SHOWS, read through the real renderer, not what its input could hold: a field the ratchet baselines is carried and dead just the same. A human node is judged per surface, because its two surfaces have different readers and each gets the whole input beside its body — the operator reads the instructions and the pause's inbound payload, the stand-in model of an llm / llm_or_human gate reads the system prompt and the whole input as its user message. Only a review gate's system prompt keeps the strict reading: that companion is handed the change under review instead (TestAReviewCompanionReadsItsInputThroughItsSystemPromptAlone). No exemption list. It reads the input.x spelling; vars.x and outputs.p.x in prose go dead the same way, and no guard holds that count — a forked session can carry such a value, so the reading would have to follow session edges.
• The schema that empties out. A node whose input schema loses its last field gets C140, "which has no field — the node's input will always be empty". Delete the schema and the node's input: line instead. An edge left with an empty with { } raises nothing at all, so remove that block by hand; and iterion fmt is what settles the trailing comma a one-key block loses.

A field a node's text reads but no incoming edge maps is the mirror defect: on every run that read delivers its surface's form of nothing (above). The compiler warns (C308) and TestCatalogInputReadsAreMappedByAnIncomingEdge holds it — the entry node aside, whose floor is the launch payload plus the vars that declare a DEFAULT, so a read of a defaulted var is answered there and a read of an undefaulted one the launch omits delivers that nothing like anywhere else (pkg/runtime/engine_resolve.go, TestBuildNodeInput_EntryFloorSeedsOnlyDefaultedVars). ir.NodeInputKeys answers what a node's input CAN hold, which on the entry node is an upper bound: every declared var, because a launch may supply any of them.

Two texts read input against something else entirely, so neither C308 nor this guard speaks for them — and a with mapping there is dead plumbing:

• a fail message: is rendered at fail time against the RUN's inputs, the launch payload (failOutcome → resolveMapping): declare the value in vars: and write {{vars.x}}, which the compiler checks. An {{input.x}} there is C149's warning, the run-inputs class;
• an interaction_prompt: is the stand-in model's system prompt when the node's own delegate asks questions, and ITS input is the QUESTION map (ExecuteHumanLLMForInteraction): {{input.<question_key>}} is the spelling that renders there, and the keys exist only at run time, so the compiler says nothing either way.

A human node's system:/model: outside interaction: llm/llm_or_human/ review — and its review-gate clauses outside interaction: review — are dropped by the compiler; that is C313, not a channel to render.

Every guard named here is part of task check.