Implement VECTOR/CUBE first-use task context after EXTCTX ABI is fixed

[zhoubot]

Scope

Implement the Linux side of the unreleased LinxISA v0.58 VECTOR/CUBE first-use contract once the architectural cross-ACR context ABI is complete.

Evidence

Audited against:

• root superproject PR #181 head 3c2ca6ef
• Linux 1055a743f16eaebfc371e0aabec8c861ab44858f / tree e6f3593f698db1de6bff8f9a43dc9007c237f21b
• LLVM b7c83f68bf84125e696a70bec4b665c70a3b584d
• QEMU PR #74 evidence head bbe090a8956bf2dce443c907e5118212d7c8768d
• r7 evidence /private/tmp/linx-pto-cube-system-181-r7/evidence

The r7 CUBE failure is not a first-use exception. QEMU observed ACR2 ECONFIG=0x7, CUBE kind 1, prepared=0, then delivered an ordinary illegal-instruction trap (trapno=0x8000000000000000, no ARGV/cause 4).

Current Linux gap

Linux 1055 defines the V/C bit constants but does not implement the contract:

• arch/linx/kernel/head.S writes zero to SSR_ECONFIG during boot.
• There are no TIF_VECTOR / TIF_CUBE used flags.
• There is no independent LIVE/resident state for NEVER_USED, SAVED_NOT_RESTORED, and LIVE.
• do_trap_insn_exception() treats E_INST/EC_PERM(4) as a page-fault-style permission exception; it does not recognize the full E=1,ARGV=1,TRAPNUM=0,CAUSE=4,TRAPARG0={0,1},BI=0 signature.
• switch_to() has no extension save/restore or per-task ECONFIG rewrite.
• fork, exec, exit, migration, and allocation-failure behavior are undefined.

V/C must remain disabled until a real handler and context ABI exist; merely enabling them would convert legal user VECTOR/CUBE headers into unserviceable traps.

Architectural prerequisites

Blocked on LinxISA issue (cross-link after creation) defining:

1. exact VECTOR and CUBE context blob size, alignment, version, reset image, and field coverage;
2. an ACR1 operation that saves/restores the managed ACR2 task context, not the currently executing ACR1 context;
3. inclusion of all task-owned state, including queues/block progress and any CUBE accumulator or extension-owned state;
4. save/restore failure, restart, migration, and forward-progress rules.

The current QEMU bring-up ESAVE/ERCOV implementation transfers a 64-byte LB/LC blob for the currently executing ACR. That is insufficient for an architecture-correct ACR1 Linux scheduler implementation.

Linux acceptance criteria

• Allocate each context only on first use.
• Track used and resident/loaded state independently for VECTOR and CUBE.
• Recognize only the complete first-use trap signature.
• Initialize or restore the selected context, clear only its V/C bit, and retry the original block header.
• On switch-out, save only used+LIVE contexts; use a defined immediate or lazy restore policy on switch-in.
• Rewrite V/C for the next task before every ACR2 return.
• Correct fork/clone, exec, exit, CPU migration, and allocation failure semantics.
• Add focused trap-envelope/state-machine tests plus a two-task QEMU runtime test proving isolation and retry.

[zhoubot]

Architectural prerequisite filed as LinxISA/linx-isa#182: LinxISA/linx-isa#182

[zhoubot]

r8 diagnostic rerun confirms this issue is not the current CUBE failure. At the exact failing header, QEMU logged LINX_FIRST_USE_TRACE with acr=2, kind=1, econfig=0x7, prepared=0; accepted the CUBE header; collected three mathematical sources and a 4 KiB FP32 output; then failed later in CUBE execution and delivered an ordinary illegal instruction. Evidence log SHA-256: a573eba8b130cced128528e41af0ec38d803ede495207974650da36ae43e26ab. Linux V/C remains safely disabled pending LinxISA/linx-isa#182.

[zhoubot]

Exact r13 sequential PTO CUBE context-switch diagnostic: children 1-2 PASS, child 3 starts but does not return within 1200s, with no user trap or explicit FAIL. Pins: pto-kernels 5f5cf061, TileOP bd1ecca9, LLVM b7c83f68, QEMU 2ba240fd, Linux 1055a74. Evidence is /private/tmp/linx-pto-cube-system-181-r13/evidence/{summary.json,qemu.log}.

A cold-boot-per-case release matrix is being added to prove each workload independently; this issue should retain the sequential cross-process Tile context/state leakage as a separate Linux scheduling/context-management blocker.

[zhoubot]

Fresh-boot split confirms workload/QEMU CUBE semantics independently: exact cold-boot matrix PASS 6/6, consistent provenance and clean shutdown for every case. Aggregate /private/tmp/linx-pto-cube-system-181-r14-matrix/aggregate_summary.json, SHA-256 8270839ea5b8f804d6cc23a429e0cc11a6edea22fa4a611241986081e69738b7.

This strengthens the classification of r13: the remaining failure is specifically sequential cross-process Tile context/state handling, not any individual CUBE workload semantic failure.