docs(isa): Sync MkDocs site to PTO-Gym v0.6 micro-instruction SPEC

[Crystal-wzy]

Align the pto-isa MkDocs site (source for https://pto-isa.github.io/)
with the authoritative PTO-Gym v0.6 PTO-micro-Instruction-SPEC. This is
a documentation-only sync; no runtime, backend, or build-script changes
beyond MkDocs nav/page-list wiring for the new and renamed pages.

Summary

• Alignment-state pages: correct pto.init_align to be store-side only
  (initializes carriers for vstus/vstur/vstar/vstas/pstu; load streams
  still start from vldas); add 9-cycle latency on vldas/vldus/vstus;
  add no-post A5 interface notes for vldus (hidden base field, reuse
  semantics); add %offset/%base "advances stream, not absolute
  store" semantics for vstus
• Conversion ops: add three missing pages — pto.vbitcast (vreg
  bitwise reinterpretation), pto.pbitcast (mask granularity bitcast),
  pto.get_vms4_sr (VMS4_SR status read after pto.vmrgsort4);
  cross-link from conversion-ops and micro-instruction landing pages
• DMA Copy section: replace the pre-v0.6 surface (pto.copy_gm_to_ubuf
  plus standalone set_loop_size_* / set_loop_stride_* configuration
  ops) with the v0.6 grouped form pto.mte_gm_ub / pto.mte_ub_gm /
  pto.mte_ub_ub using inline nburst(...) / loop(...) / pad(...)
  clauses; add new pto.mte_ub_l1 page; mark the six standalone
  set_loop_* pages deprecated with migration banners; URL slugs
  preserved to avoid breaking external links
• New /isa/cube/ section (previously absent): README + three
  architectural pages (NZ Fractal Layout, Buffer Hierarchy, FIXPIPE
  Model); 6 MAD ops (mad, mad_acc, mad_bias, mad_mx,
  mad_mx_acc, mad_mx_bias) with shared MAD Common Clauses table
  and MX Matmul Model; 13 data-movement ops (mte_gm_l1,
  mte_gm_l1_frac, mte_l1_ub, mte_ub_l1, mte_l1_l0a/l0b,
  mte_l1_l0a_mx/l0b_mx, mte_l1_bt, mte_l1_fb,
  mte_l0c_l1/gm/ub); each page follows the existing pto-isa
  per-op template
• Predicate interleave/deinterleave variants: add the four missing
  granularity variants — pto.pintlv_b8, pto.pintlv_b32,
  pto.pdintlv_b16, pto.pdintlv_b32 — so each family covers
  b8/b16/b32 per the SPEC; also retype the existing pintlv_b16 /
  pdintlv_b8 syntax to explicit !pto.mask<bN> and refresh the
  predicate-generation-and-algebra landing pages to list all six
• Dual load/store naming: rename pto.vldx2 → pto.vldsx2 and
  pto.vstx2 → pto.vstsx2 to match SPEC v0.6 across page titles,
  syntax / assembly forms, and landing-page entries; rename the per-op
  files and update all 21 cross-references including the mkdocs.yml
  nav and gen_pages.py path list; add 9-cycle (vldsx2) and
  12-cycle (vstsx2) A5 latency entries sourced from SPEC §III
• SFU/DSA naming: rename pto.vexpdiff → pto.vexpdif (single 'f',
  SPEC v0.6); page renamed + 22 cross-references updated across docs
• Predicate granularity (pset / pge / plt _bN): correct syntax
  to typed !pto.mask<bN> instead of bare !pto.mask for the 9
  instructions (b8/b16/b32 each); English + _zh updated
• !pto.mask granularity sweep (~800 docs): replace every remaining
  bare !pto.mask reference with the SPEC-canonical parametric form
  !pto.mask<G>, then upgrade <G> to concrete <b32>/<b16>/<b8>
  in worked-example lines whose surrounding !pto.vreg<NxT> carries a
  concrete element family. Template lines and mixed-family lines keep
  <G> (parametric). 0 bare-mask occurrences remain after the sweep
  (782 → 0)
• New /isa/tile/view-and-tile-buf section (previously absent): SPEC
  Tile §2.4-2.5 + §3 define foundational tensor-view and tile-buffer
  descriptor/handle operations that the tile compute surface depends
  on. Add 9 op pages — pto.make_tensor_view,
  pto.get_tensor_view_dim, pto.get_tensor_view_stride,
  pto.tensor_view_addr, pto.partition_view, pto.alloc_tile,
  pto.subset, pto.set_validshape, pto.tile_buf_addr (the
  tile↔vector bridge inside pto.vecscope) — plus a landing page and
  tile README / current-isa-scope / mkdocs nav wiring
• Internal H1 normalization: change docs/isa/system/ops/TFREE.md H1
  from # TFREE to # pto.tfree for consistency with TPOP/TPUSH;
  normalize 11 _zh comm-section H1s (TBROADCAST_zh, TGATHER_zh,
  TGET_zh, TGET_ASYNC_zh, TNOTIFY_zh, TPUT_zh, TPUT_ASYNC_zh,
  TREDUCE_zh, TSCATTER_zh, TTEST_zh, TWAIT_zh) from uppercase
  TUPPERCASE form to # pto.t* form
• Strip UTF-8 BOM (U+FEFF) from 28 affected docs/isa/**/*.md files
  for consistent file encoding
• Chinese (_zh) translations for every new and updated page, kept
  structurally identical to the English counterparts
• MkDocs wiring: register the renamed vldsx2 / vstsx2 / vexpdif
  files and the new cube / view-and-tile-buf nav subtrees in
  docs/mkdocs/mkdocs.yml and docs/mkdocs/gen_pages.py

Testing

• cmake --build build/docs --target pto_docs builds cleanly, no
  WARNING from MkDocs nav resolution
• All updated and new pages render under build/docs/site/
• Spot-checked rendered HTML for v0.6 syntax, 9-cycle latency
  notes, deprecation banners, cube section navigation, the renamed
  pto.vldsx2 / pto.vstsx2 titles, the new view-and-tile-buf
  section, and the normalized comm/system H1s
• _zh files mirror English counterparts structurally

[gemini-code-assist]

Code Review

This pull request provides a major update to the PTO ISA documentation, introducing detailed specifications for Cube micro-instructions, the FIXPIPE writeback model, and NZ fractal layouts. It also updates the scalar DMA copy documentation to reflect the v0.6 grouped transfer interface and adds references for new conversion and status query operations. The review feedback pointed out a consistency error in the L0B physical layout description and identified several missing type parameters in the MLIR syntax and examples for the pto.mte_gm_ub instruction.

[zhoubot]

Reviewed the updated PR against current origin/main after PR 131.

No blocking findings from this pass. Verification:

• Merged origin/pr/126 into a detached current-main worktree; merge was clean.
• git diff --check origin/main...origin/pr/126 passed.
• mkdocs build -f docs/mkdocs/mkdocs.yml --strict passed locally.
• Checked the previously reported CI failure: the failing GitHub full-ST job was the stale cpu_stub.hpp / SyncCoreType compile error that PR 131 fixed on main. In the merged local worktree, the CPU ST and CPU comm ST builds progressed past that failure. The local full runner then stopped during data generation because this machine lacks Python package ml_dtypes for tmatmul_mx, so I could not complete the full runtime suite locally.