fix(vpto): close Issue #24 lowering gaps for PyPTO merged-device A5 kernels

[liuzidi]

Fix Issue #24: VPTO backend lowering failures for PyPTO merged-device A5 kernels

Fixes #24

Summary

On the PyPTO merged-device VPTO route (--pto-backend=vpto --pto-level=level3 --vpto-emit-merged-device-only, one kernel per compilation unit), 26 of the 27 DeepSeek-V4 Pro operator test cases (pypto-lib/models/deepseek_v4_pro/*.py, platform=a5) failed at PTOAS compile time. This PR closes all six compile-time failure classes from the issue analysis.

Result: 16 of 27 cases now compile end-to-end (previously 1). The remaining 11 are blocked on the tpush/tpop/tstore VPTO data-plane lowering (loop-carried fixpipe slot rotation), which is a separate architectural gap documented as a follow-up below — everything up to that boundary now works.

Root causes fixed

#	Root cause	Fix
1	pto.trsqrt 3-operand form (src, tmp) → dst had no TileLib template (11 cases)	template_trsqrt_tmp / template_trsqrt_tmp_1d via a new register_unary(has_tmp=True) flag; vmi_trsqrt_with_tmp_narrow for compact rows below the 128-byte elementwise minimum (UB slots are planned 256-byte aligned, so a full-row VMI access stays in-slot)
2	pto.tgather 4-operand index form had no template; _axis_is_row/_axis_is_col param-name bug made every mask-form candidate reject; packed-sort-key gather needs cross-dtype signatures (20 cases)	template_tgather_tmp (tmp not read by the A5 vgather2 sequence; tmp dtype == indices dtype; index width follows data width, mirroring the op verifier); fixed the axis param names; equal-storage-width cross-dtype mask signatures with pto.vbitcast on the result; InsertTemplateAttributes exports axis/cmp_mode/offset context attrs
3	pto.tmax / tcolexpand / trowexpand / trowexpandmul constraints rejected real PyPTO tile shapes (2 cases)	template_tmax_dst_valid_prefix (src1 valid cols strictly less than dst — strict so it stays mutually exclusive with template_tmax and selection stays unambiguous); _valid_column_expand accepts src valid cols ≤ dst valid cols; trowexpand accepts [rows,1] col-major single-column sources; expand-binary valid-row check eq → ge
4	pto.tsetval / pto.tgetval had no VPTO lowering → unrealized_conversion_cast (11 cases)	FoldTileBufIntrinsics rewrites tile scalar IO to tile_buf_addr + store_scalar/load_scalar before addr folding (VEC space + element-type match enforced; hard error otherwise)
5	pto.initialize_l2l_pipe peer-init pairing failed when the peer kernel lives in another compilation unit (4 cases)	PTOResolveReservedBuffers accepts incomplete peer groups when all inits are cross-module pipes, records unresolvable imports under a synthetic peer key, and materializes the import base to the shared pipe-contract constant (PyPTO plans both sides from base=0). The relaxation is gated on the single-kernel-per-file shape (isSingleKernelVptoUnit: exactly one authored, non-private func.func in the unit), so multi-function units still reject bad peer_func symbols — the three existing guard tests keep passing
6	One-entry-per-ELF vs AIC+AIV mixed groups (13 cases)	VPTOSplitCVModule stamps every entry candidate when each kernel kind contributes at most one candidate; ImportReservedBufferOp::verify consults a driver-installed parse-time backend hint so single-kernel units parse before the module attribute is stamped

VMI SPEC compliance

The new VMI template (vmi_trsqrt_with_tmp_narrow) and the VMI-related constraint helpers follow the VMI spec: the pto.vmi layer only emits vector ops on logical vector/mask values (scalar-pipe ops stay in the pto.mi layer); requires_full_physical_row=False is only used where memory planning guarantees the slot is padded to 256 bytes, so a full-row access cannot over-read the slot.

Testing

• lit (pto + vpto + vmi_new): 1874/1874 pass, including the three pre-existing peer-lookup guard tests and the new regression test test/lit/pto/import_reserved_buffer_single_kernel_unit_resolves_a5.pto (verifies: parse-time peer check relaxed for the single-kernel shape → resolve pass materializes the import to the pipe-contract constant → pipeline proceeds past peer resolution into VPTO LLVM emission, failing only at the documented tpop data-plane gap).
• ptodsl unit tests: tilelib suites 144/144 pass; all 9 test_vmi_*.py scripts pass.
• PyPTO DeepSeek-V4 Pro 27-case sweep: 16 compile (was 1). Verified no constraint/priority regressions: the tmax strict-prefix fix removes a selection ambiguity, the trsqrt _tmp templates use distinct candidate IDs, and the tgather _tmp signature mirrors the op-verifier dtype coupling (f32→i32, f16/bf16→i16).

Known follow-up (out of scope)

The remaining 13 cases stop at the tpush/tpop (fixpipe FIFO slot rotation, loop-carried tile index) and one tstore constraint-form VPTO data-plane lowering gap. That requires lowering tpop to TASSIGN at CONSUMER_BUF + (tileIndex % SLOT_NUM) * SLOT_SIZE with loop-carried slot rotation — large enough to warrant its own PR. This PR unblocks everything up to that boundary (peer resolution, template selection, scalar IO, entry stamping all work).

[liuzidi]

独立 review 记录（GLM-5.3，独立新会话）：

• Review Merge pull request #8 from liuzidi/feature-vmi-shengtao #1（commit 49fb751，六类根因修复）：VERDICT: PASS，4 个 P2 意见（narrow 约束上界、trowexpand dst 收紧、pipe 回退 remark、tmax 去重）。review 会话 transcript 132/132 条 model_request 均为 GLM-5.3。
• Review [Bug] VMI tile-template fallback path 把非物理倍数列数当 mask size，撞 VMI_LANE_COUNTS 校验（如 1x32 f32 texpands） #2（commit e35a3cb，P2 修复收敛）：VERDICT: PASS，4 个 P2 全部确认"resolved correctly"（含 128B 边界 exhaustive & disjoint 的运行时验证、verifyTRowExpandCommon 可达集比对、remark 生命周期分析、循环导入检查）；仅 2 条 informational P3（未使用的对称参数、单向交叉引用注释），reviewer 明确"no action needed"。review 会话 transcript 21/21 条 model_request 均为 GLM-5.3。

两次 review 的所有意见（4×P2）均已修复并经 1874/1874 lit、144/144 tilelib 单测、9/9 VMI 脚本回归验证。

[github-actions]

Warning: @liuzidi, ci-sim exceeded its soft runtime budget.

• vpto-sim-validation runtime: 24h 0m 0s
• Soft budget: 1h 30m
• Job conclusion: cancelled
• Workflow run

This warning is advisory only and does not affect required checks. Please inspect the step timings for an unexpected regression.