S0 servicer extraction review - from the bus-interface session¶

Full path: E:\Dev\Ronny\NDInsight\SINTRAN\ND500\S0-SERVICER-REVIEW-FROM-BUS-SESSION-2026-07-17.md Reviewing: RetroCore commit 82e83a148 (frozen contract + S0 extraction). Reviewer basis: the live traces (3 runs), the 030-S3SM5/nd-500-mon carves, the B30 microcode pseudocode, and the swapper startup analysis (catalog sections 7c/7d).

Verdict¶

S0 extraction verified faithful. Checked line-by-line against the pre-extraction engine (f79937857): N5STA lifecycle + power-fail-bit preservation, MICFU cases (13B/14B 32-bit addressing, 12B no-op, 22B accept, known-set accept, unknown->5ERANSWER), event ordering, log formats ("@word" = msgBase>>1 == old mar&0xFFFFFF), mar==0 bare-activate guard kept in the host, MAR word->byte <<1 conversion at the host boundary, DEBUG_DETAIL hardcoded before AND after (no change). IServicerHost.Nd500AddressBase == SharedMemoryStart, AnswerWritten no-op on classic, completion stays with ExecuteND500Operation - all as documented. Full ND-500 suite re-run in progress (architect reported 1751/0; will confirm).

Findings (ranked)¶

F1 - GENERATION POLICY HAZARD (design-level, hits the live oracle in S1)¶

Nd500Generation doc says semantics follow the LOADED MICROCODE, and the live machine runs a 5800-revision image behind the classic 3022. But on a true 5800 table, MICFU 21B (3WREG) and 20B (3RREG) are MSG_ILLEG - and the live SINTRAN L LOAD-SWAPPER path REQUIRES 21B accepted (it is the current blocker; catalog 7c). If S1's generation table strictly follows the loaded image, the live oracle regresses the moment anything switches the servicer to Samson5800 (e.g. "detect image revision"). Recommendation: the effective authority for accept/reject is the SINTRAN SENDER, not the image; keep the 3022 host's servicer on Classic regardless of loaded image, and never auto-derive Generation from CS content. The B30 "21B=ILLEG" belongs in the Samson5800 column only.

F2 - CONTRACT GAP for the live blocker: 21B 3WREG needs a servicer->CPU deposit path¶

21B = REGISTER WRITE (process-0 initial register context incl. P; layout: first-reg@7, count@10B, 32-bit ND-100 src addr@11B-12B; answer = status only - catalog 7c, Micro Program Guide 13.13). Neither sink covers this direction (both are CPU->servicer). Needed: an IND500Cpu extension (e.g. WriteRegisterBlock(firstReg, values)) or a servicer-side context stash applied at CPU attach. Twin 20B 3RREG needs the reverse (register values -> ND-100 memory). Without 21B the live oracle cannot pass LOAD-SWAPPER, so this belongs in S1/M1, not later. (Register-image width in ND-100 memory - 2 words/reg hi-first - is INFERRED; next live run should log offsets 10B-12B.)

F3 - CONTRACT GAP for S1's 3RMICV: no version source on IServicerHost¶

The 3RMICV two-halfword answer (version + CPU param, landing at HW 7 + 10B) must source the version FROM THE LOADED CS IMAGE (acceptance oracle: classic images self-report 10509/10609 via READ MICRO PROGRAM VERSION; the 5800 image = 027232B). csStore lives in the 3022 host; the servicer has no way to read it. Suggest ushort MicrocodeVersion { get; } (or a version+cpuParam pair) on IServicerHost, host-updated at CS load/restart. Note the classic 144-bit image's version LOCATION inside the image is still UNKNOWN - extract empirically when Bo Goran's floppy arrives (search 10509=0x290D / 10609=0x2971).

F4 - DOC BUG in Nd500MicrocodeServicer.MessageProcessed XML comment¶

Comment says the event fires "after the answer status is written"; the code fires it BEFORE (invoke precedes the N5STA write). The CODE order is the correct one - result data must be in place before N5STA:=3 releases SINTRAN (matters for M1's answer payloads). Fix the comment, keep the code.

F5 - Threading note (not an S0 defect)¶

The classic host still calls servicer.ProcessMessage synchronously on the ND-100 IOX thread (ExecuteND500Operation). Fine under the contract's inline mode; when the CPU thread arrives, the classic activate must be enqueued through the SPSC ring too, or the servicer runs on two threads. Flagging ExecuteND500Operation as the spot that changes.

F6 - In-flight work / naming migration: clean¶

Nothing in the bus session assumed the nested enum location or the private engine body - everything was committed at f79937857 and the three migrated test files compile against the new names. Bus-session docs (skill + NDInsight) referenced "the engine in NDBusND500IF" - being updated to point at the Servicer namespace (doc-only, no code).

Live-oracle status¶

Pending a fresh EXE build + status run on the live machine (needs Ronny). Expected decoded-trace equivalence: identical MAILBOX RECV/ANSW lines incl. the 21B 5ERANSWER (still unimplemented by design in S0). Will report after the run.