Fabric switch reader: RouterOS response → typed lane readings

[briansrls]

First code increment of the modeled switch-convergence path (design in #11524). This is the read path's functional core: it turns a RouterOS /rest/interface/ethernet response into the Crs812LaneReading values gunbc.spark.fabric_switch_assessment was written to receive. Today fabric_switch_subject() passes readings: [], so every lane resolves to unknown; this is the producer that closes that gap on the read side.

What's here

• extdeps.mikrotik.crs812 — the inverse wire parsers (DESIGN §4, one grammar both directions): crs812_lane_speed_from_wire, crs812_fec_mode_from_wire, crs812_wire_bool, and the Crs812EthernetWire record. Each is partial and fail-closed: a wire value the corpus doesn't model decodes to an explicit Unmodelled arm carrying the string, never a silent default.
• gunbc.spark.fabric_switch_read — the workflow-layer decode: filters the GET response to our fabric legs (by the RouterOS comment the switch carries), and turns each per-lane wire record into either a Crs812LaneReading or a located refusal naming the interface, field, and raw wire that didn't parse.
• Witness — 6 claims, green by execution against the real response captured from gunbc-fabric-switch (192.168.1.240) on 2026-09-17: the 8 legs decode to 50G/fec91/enabled, a non-fabric interface is filtered, an unmodelled speed refuses at speed, an unmodelled FEC refuses at fec-mode, a disabled lane reads not-enabled.

Scope boundaries (deliberate)

• No live transport. Fetching the JSON over the switch's basic-auth REST is blocked on the same capability extdeps.bmc.http declared as debt (redfish_http_hardwired_transport_dissolution_trigger: the shared transport rest machinery doesn't yet realize transport-level basic auth). So this is the pure decode — wire records in, typed readings out — fully testable with no network. The live fetch lands beside it on the BMC precedent. A pure decode is also the right seam: the same function serves a live read and a replayed fixture without branching.
• No autoneg axis yet. The reading type stays as-is (speed/fec/enabled); adding the autoneg dimension (gap G1) is a separate change to the load-bearing assessment type.
• No converge apply. Deferred until the hardware links at 100G (§4d — its FEC/autoneg specifics are inferred, not witnessed).

🤖 Generated with Claude Code

[gunbai-bot]

All four findings in review 67287 addressed in the latest commit; 9 claims green by execution.

1. §3c dangling reader — fixed by landing the join. You're exactly right: the stated BMC-transport frontier blocks the fetch, not the join, and turning LaneReadOk outcomes into the List<Crs812LaneReading> that crs812_switch_subject already takes is pure and network-free. Landed that chain:

• crs812_fabric_read — the join, fail-closed at the read grain (one unparseable lane refuses the whole read rather than handing the assessment seven readings where the switch reported eight).
• fabric_switch_subject_from_wire — builds the real Crs812SwitchSubject that inspect_fabric_switch consumes.

So crs812_fabric_lane_read_outcomes is now consumed by the join, the join by the subject-builder, and the only declared frontier is that terminal function — whose trigger genuinely is the live fetch (the BMC basic-auth transport debt), stated on it.

2. §3c unread fields — dropped. auto_negotiation and running are gone from Crs812EthernetWire; no fold read them. They return with their consumers (autoneg axis G1, live link-state G3), not before.

3. §4b rung honesty — comment fixed. You caught the real thing: the fixture is hand-authored (transcribed from the 2026-09-17 response), so "executing over the captured bytes" was rung inflation. Reworded — the records are supplied inputs, the claim discriminates the decode not the fetch, and the inhabitance pairing (a real JSON→wire producer emitting this shape) lands with that fetch.

4. Predicate-for-a-test (advisory) — resolved by #1. crs812_lane_read_outcome_refused now has a production consumer: crs812_fabric_read folds it.

New discriminating claims: the join produces eight readings for the subject, one bad lane refuses the whole read, and the frontier builds a subject from a clean read. Thanks — the dangling-reader call was the right one; the join was pure all along and I'd left it on the wrong side of the frontier.