fix: refuse a deck with no frequency, instead of succeeding in silence

[dc0sk]

The defect

fnec deck.nec on a deck with no FR card exited 0 having written zero bytes to stdout and stderr. A silent success is indistinguishable from a run that worked — and it was one of four different answers to the same deck (FND-070):

frontend	before
CLI	exit 0, nothing at all
GUI (two seams)	refused — "deck has no FR card"
fnec_py::solve_deck_str	raised — "deck has no FR card"
fnec_py::sweep_deck_str	returned [], at success
CLI sweep --resonance	refused — "resonance search: deck must have an FR card"

That fourth row my own site enumeration missed; the design review found it. It is the defect inside the frontend FND-070 credits with refusing — one module disagreeing with itself, its sweep half returning the empty-result-standing-for-an-error shape that docs/json-output-schema.md now tells consumers not to read that way.

Why the predicate is typed on the resolved list

validate::no_frequency_error(freqs_hz, remedy) takes the resolved frequency list, not the deck. That is the load-bearing choice, not a detail.

Frequencies do not have to come from an FR card: --sweep-config supplies them for the CLI, the sweep tab's own range widgets for the GUI, the wire for the worker. So this cannot live in pre_solve_error — the gate every frontend already calls, and the obvious home — because that sees only the deck and would refuse a working, documented capability.

$ fnec --sweep-config sweep.toml no-fr.nec     # measured: exit 0, both points solved

No test covered that capability until this change added one. docs/cli-guide.md said --sweep-config "overrides the FR card", which is true and incomplete; it also supplies one.

Sabotage — this is the improvement over FND-145

Forcing the helper to return None:

gate	result
CLI sweep_contract	FAILED
GUI gui_smoke	FAILED
pytest, freshly built wheel	FAILED
controls (--sweep-config solve, FR-deck sweep)	green

FND-145's equivalent sabotage left every CLI test passing, because the CLI had a second route to that refusal. Here there is none, so every frontend notices — which is what sharing the predicate buys over sharing only the sentence.

FND-084 is closed by the opposite of the fix it proposed

That row asked for "one line: emit [] before the early return". Rejected, for reasons the review supplied and I had not considered:

• [] is printed only under --output-format json, so text mode would have stayed silent and FND-070 would have survived intact — it fixes the symptom, not the defect;
• [] + exit 0 already means solved, and there was no feedpoint to price (measured on corpus/dipole-ex1-freesp-51seg.nec). Spending it here makes "never solved" indistinguishable from a real result — in the very document where, one PR ago, I wrote that an empty array had been silently standing for an error and that consumers must branch on the exit code;
• there is no contract to honour: json.loads('') raises, so every consumer in the tree already crashed on this deck and now gets a clean non-zero exit instead.

FND-070 and FND-084 have sat adjacent in the ledger since the 2026-08-28 audit proposing opposite fixes, and nobody noticed. Closing either as written would have entrenched the other.

Breaking

Breaking for anyone scripting fnec on a frequency-less deck and reading exit 0. Deliberately divergent from nec2c, which defaults such a deck to 299.8 MHz (λ = 1 m) and answers it — on a 10.5 m dipole it reports 133.18 + j280.36 Ω, pricing the wire as 10.5 λ. A plausible number for a deck the user did not write.

This is the second breaking change to this contract paragraph, and it follows the maintainer's "refuse everywhere" decision on the identical no-EX case rather than a fresh instruction — flagged rather than assumed.

Recorded, not fixed

• FND-150 — a bad --hosts path was never reported on this deck, because the early return preceded reading the file.
• FND-151 — the frequency checks now live at two seams plus the worker's wire check; folding them into a pre_solve_error that takes the resolved list is a four-frontend signature change, deferred deliberately.

Gate

scripts/check-all.sh → EXIT=0, 15/15, on the committed tree. Python gates run against a freshly built wheel: 30 pytest tests pass, and both new ones fail under the sabotage.

🤖 Generated with Claude Code

https://claude.ai/code/session_018p7FxX7QMWNJVNp9LbaLkB