[mrmidi]

Config ROM cache and scanner fixes around bus resets and the 8-bit bus generation wrap.

The bus generation is the raw 8-bit Self-ID value and rolls over from 255 to 0 (OHCI 1.2 Table 11-2), so generation 0 is a real value and generations can only be compared for equality. Linux (ohci.c:2143, -1 as the invalid sentinel) and Apple (IOFireWireController.cpp:2750, scan->generation == fBusGeneration) both do it that way. Several places here used Generation{0} as "none" or ordered generations with < / >.

Builds on gly11's #87 (the first four commits are cherry-picked with -x, authorship kept).

Cache lifecycle across bus resets

• Suspend cached ROMs on bus reset; selector 14 refuses suspended entries (gly11, fix(config-rom): reject suspended generation exports #87). ExportConfigROM is now exact generation or nothing.
• Drop the unreachable ValidateROM path (gly11, fix(config-rom): reject suspended generation exports #87).
• Lookups return copies taken under the lock. FindByNode returned pointers into the store's maps after unlocking, while the driver queue inserts/erases and the user-client queue reads.
• Drop ROMs suspended by the previous reset. The (generation, node) index grew by one entry per node per reset, and after a wrap a reused key could carry another device's ROM.
• One FindByNode(gen, node, allowSuspended) with no default, so every caller states whether it accepts a pre-reset ROM.

Generation 0 and the wrap

• Abort the scan on every reset, including when the previous generation was 0; Abort/IsBusyFor no longer treat 0 as "no scan".
• Scans at generation 0 complete. ROMScanSession used gen_ == 0 as "not running", so after the 256th reset MaybeFinish never fired and the IRM verify replies were dropped. It now keeps an explicit scanActive_.
• IsBusyFor uses the generation recorded at Start. The session only learns its generation when its start task runs, so a generation-0 request could be refused as a duplicate in between.
• GUID cache and unread lookups removed. The GUID index and FindLatestForNode ordered generations with < / >, which is backwards after a wrap. Nothing read either: the export fallback was dropped in 805def0, and FindByGuid, Snapshot, SnapshotByState, Clear had no callers. The store is now a single (generation, node) index; the manual ROM re-read uses EraseByGuid.
• Removed firstSeen / lastValidated (unread, and filled in using 0 as "unset") and the unused GenerationContextPolicy.

Reuse unchanged ROMs after a reset

• A node whose bus info block is unchanged completes from the previous generation's ROM. Discovery read every node's whole Config ROM again after each reset (bus info block, root directory, text leaves, unit directories), although the previous copy was in the store. On an LS-9000 that rescan was ~550 ms of the 650 ms between a reset and the SBP-2 reconnect (SBP-2: re-login/reconnect after rediscovery, not on a 100 ms timer #216), and every device on the bus waits for it. Apple reads only the 5-quadlet bus info block and keeps its cached ROM when the block is identical (IOFireWireROMCache::hasROMChanged, IOFireWireROMCache.cpp:240-320; IOFireWireDevice.cpp:727-743). The block carries the GUID, the ROM CRC and the ROM generation a device advances when its ROM changes.
• The scan gets the ROMs the reset suspended (ConfigROMStore::SuspendedROMs). After the bus info block, the speed probe and the IRM check (both still run: they are bus state, not ROM content), a matching node completes from its copy with the new generation and node ID (UnchangedROMPolicy).
• Always read in full: any difference in the block; a generation-0 ROM without units (Apple's rule, for devices that publish units late); a copy built around a failed or short read (readIncomplete), so a lost read heals at the next reset; a manual re-read.
• Only one generation back is kept (the store bound above): a reset that interrupts a scan leaves nothing to compare, and the next scan reads in full.

App

• ROM Explorer drops a selection from a stale topology. After a reset it compared nothing, polled the old generation until timeout, and could re-select the same node ID on a different device.

Testing

• C++ suite: 3819 pass, 6 skipped (same skips as main).
• New tests: generation-0 scan completion with and without the IRM check, generation-0 scan with no remote nodes, busy check before the start task runs (manual-dispatch queue), reset after a generation-0 topology aborts the scan, 255 → 0 store rollover, EraseByGuid, concurrent lookups vs SuspendAll/EraseByGuid, ROM Explorer staleness checks.
• ROM reuse: a device scanned twice reads only q0 and q2-q4 the second time; a new node number still matches; advanced ROM generation, new header CRC, another GUID, a copy with a failed read, generation 0 without units, and no previous ROMs read in full; the 255 -> 0 wrap reuses; the IRM check and speed probe still run; the scan request carries the suspended ROMs.
• Every fix was mutation-checked: restoring the old code makes a new test fail.
• Dext compiles in Xcode for both architectures. Hardware check of normal discovery / ROM export / manual re-read: pending. For the ROM reuse: after a reset with devices attached, each logs bus info block unchanged once (/usr/bin/log show --predicate 'eventMessage CONTAINS "bus info block unchanged"').

🤖 Generated with Claude Code