[gdluck]

Requested by A

Summary

lookup_firmware_verdict currently explains every NACK as "Gateway refused/rejected the command", and compare_gateways groups NACK + OUT as a separate outcome class. On F460 020012, F461 020011 and MyHomeServer1 028206 that reading is wrong for most rows: those firmwares report success only after a device TX-ACK, which the empty emulated bus never gives, so NACK with verdict OUT means "the frame went out, no device answered", not a refusal (own-firmware-oracle findings/empty-bus-reply-semantics.md; 228 of the 309 OUT rows of each of those gateways are NACK). Since schema 1.2.0 the index says this per gateway in gateways[].reply_semantics (empty_bus_txack / pic_ack / unknown); MyHOME already reads verdicts that way (MyHOME#766). An AI agent reading the current tool output concludes that e.g. *#1*06## is refused by the F461, which it is not (a live F461 answers it).

Changes (src/openwebnet_mcp/firmware_oracle.py)

• load() keeps a product -> reply_semantics map (empty string for an index without the field).
• reply_meaning(product, reply, verdict): one sentence per row. NACK + OUT on an empty_bus_txack gateway: "Frame went out on the bus; the NACK is this firmware's empty-bus rule (no device TX-ACK on the emulated bus), not a refusal." NACK + SILENT: "Gateway refused the frame (*#*0##); nothing reached the bus." Other replies keep their wording.
• format_verdict_markdown (the Ground Truth lines of lookup_firmware_verdict) uses it.
• compare_gateways(frame=…) adds a note under the outcome classes when such rows are present: NACK + OUT on those gateways is not a divergence from ACK + OUT.
• The fleet summary table gets a Reply Semantics column; status() exposes the map.
• An index without the field (the bundled 1.1.0 snapshot until chore(oracle): sync the bundled oracle index to own-firmware-oracle 53f9514 (schema 1.2.0, 529 inputs) #20 lands) keeps the plain reading: no behaviour change until the field is present.

Tests

tests/test_firmware_oracle.py: five new tests on a schema-1.2.0 index built in tmp_path (F461 empty_bus_txack, MH200N pic_ack, a catalogued F455): the map is loaded and exposed; reply_meaning for every reply/verdict pair; the Ground Truth text for an empty-bus row, a refused row and a pic_ack NACK; the compare_gateways note and the fleet column; and an index without the field. One existing assertion changed with the wording (refused/rejected → refused the frame … nothing reached the bus).

Verification

• pytest -q: 167 passed, 4 skipped; coverage 96 % (--cov-fail-under=95).
• ruff check src tests: clean.
• Works with both the bundled 1.1.0 index and the 1.2.0 one of chore(oracle): sync the bundled oracle index to own-firmware-oracle 53f9514 (schema 1.2.0, 529 inputs) #20 (sibling checkout of own-firmware-oracle at 53f9514).

🤖 Generated with Claude Code

[gdluck]

Heads-up: this PR shares tests/test_firmware_oracle.py with #22 and will conflict once that lands; I will rebase it on #22 right after the merge (the reply_semantics reading here is not part of #22, so both are needed).