fix(BACKEND-TENSTORRENT-QWEN35): EnsureDevice2D consumed stale host bytes under host-free decode; Qwen3.5-0.8B sacred gate 16/16 on ambient - #1768
Draft
lu-zero wants to merge 16 commits into
Conversation
…nt — allow-list, op delta, first e2e gate (mudler#1715) The wiring row the GDN spec's ## Owed names: register Qwen3_5ForConditionalGeneration on the TT allow-list (landing last — the discrete P150 refuses an op miss by name, never falls back to CPU), complete the op delta the 0.8B forward refuses by name (candidates from source inspection: kGdnPostConv and kSigmoidGateBf16 are direct dispatches with no composite fallback; kAttnQkNormRopeGate and kAttnQkNormRope are registration-guarded fused preambles whose unfused path needs an also-unregistered AttnGateSplit — a W0 refusal sweep pins the authoritative list), and prove the family end to end with the device-golden pair treatment the Mistral row ratified (VT_DUMP_IDS bootstrap, transformers teacher-forced near-tie gaps via scripts/qwen3-neartie-gap-transformers.py, 16/16 with BACKEND PROOF). Capacity decision recorded, not deferred: Qwen/Qwen3.5-0.8B bf16 ~1.6 GB fits the proven 14.5 GB envelope; 27B bf16 ~53.8 GB and the GGUF k-quant arms (no TT kernels) are refused by name and owed. Weights download authorized 2026-08-23; revision pinned at 2fc0636471. W3 picks up the GDN row's reviewer leftovers (d2h counter completeness, the conv_transposed fast-path check). Records: the backend-matrix row, the claim, and both checker re-pins (check-agent-record BACKEND 86->87; check-gate-commands RUNNABLE_BASELINE +row, inherited credit in the mudler#1541 shape — the full TT suite, CPU gate, and preflight run today; the row's own sweep/e2e gates are owed). Nothing implemented yet; W0-W3 owed in the spec. Issue mudler#1715 stays open until the family runs. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:zai-glm-5.3 [Maki]
…fused ops + ScatterRowsExact fix (mudler#1715) The W0 refusal sweep ran the Qwen3.5-0.8B forward on the P150 with scratch CPU stubs and surfaced three by-name op refusals this change implements as native Tenstorrent kernels (SigmoidGateBf16, GdnPostConv, AttnQkNormRopeGate), each red-first and gated doctest-vs-CPU-oracle. The softplus inside GdnPostConv composes relu(x)+log1p(exp(-|x|)) in f32 because the one-shot SFPU softplus poly carries ~1e-6 absolute fit error that becomes 4.4e-4 relative on small g values (measured, first green attempt failed 5/100). The q|gate split of AttnQkNormRopeGate happens in the host gather because slice+reshape on a fresh from_vector TILE upload returned wrong data on the P150 while the full-width leg stayed correct; every leg now uploads already-shaped. ScatterRowsExact moves to a zero-copy ROW_MAJOR view + indexed_fill so the Qwen3.5 GDN cache shapes stop OOMing (32 GiB TILE inflation) and overflowing L1 (ROW_MAJOR reshape); kGdnDecode/kGdnStateScatter register the real kernels, and all five W0 sweep stubs are deleted so the tree carries only real kernels. The scratch arch allow-list is reverted; the arch stays refused at load until W2, and the CausalConv1dUpdate bf16-state capability refusal is escalated in ## Owed. Full gate both legs (ambient and VT_TT_HOST_FREE_DECODE=0): 36/36 cases, 2259/2259 assertions. Negative mutations M1-M4 all red, restored byte-for-byte. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:zai-glm-5.3 [Maki]
…ow-list, split-shadow scatter, pool-tenancy seam, gemma bake (mudler#1715) Lands the W2a and W2b engine waves for the Qwen3.5 GDN family on the Blackhole P150, each pinned by the evidence recorded in .agents/specs/tenstorrent-qwen35.md. W2a: SupportsCompressedConvState/SupportsCompressedGdnState flip to true (the TT kernels compute through f32 shadows and re-round committed values through bf16 storage, mirroring CUDA), the arch joins TenstorrentPlatform::supports_model_architecture last in implementation order, ScatterRowsExact stages born-split shadows (SplitFactor keeps indexed_fill pages inside L1 — the 262144-wide ssm_state row previously threw "grow to 2208704 B"), and the parity gate gains its Tenstorrent device-golden harness plus wide-row gather/scatter/decode arms. Pool tenancy: DevicePool::Get hands recycled blocks to new tensors without passing through Alloc, so Tenstorrent's host-pointer-keyed device shadows survived tenancy changes and the next reader downloaded the PREVIOUS tenant's dead tensor (got 2048 want 4096 on a [4,1024] readback). New vt::Backend::OnScratchBlockAcquired (default no-op) is called on every free-list hit; Tenstorrent overrides with MarkHostWritten. CommitDeviceLogical2D now asserts the committed device volume against rows*cols so producer shape bugs die at their producer. Gemma: RmsNormKernel's device arm handed ttnn::rms_norm the raw affine, but Qwen3.5 sets gemma=true at 21 sites — under host-free decode every forced- device norm dropped the +1 and collapsed prefill to degenerate tokens. The device arm now bakes w+1 host-side in f32 (the oracle's order, as the fused preamble's weff does) into a transient gamma that bypasses the raw- affine slot cache. Pinned by a new op-level case whose drop-the-bake mutation goes red. Diagnostics (env-gated or error-path-only): DownloadToHost reports caller context, device shape/dtype/layout and a backtrace on mismatch; VT_TT_SLOT_TRACE logs slot register/unregister/commit/ensure; the d2d copy hook reports pointers and frames. All four located today's defects. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
… diagnosis chains, fixed-engine golden pair (mudler#1715) The spec's W2b sections record the full bisection chains and evidence logs: the pool-tenancy defect (slot-trace pinned the stale [8,256] bf16 shadow to the previous tenant of a recycled DevicePool block; mutation cell closed at e2e depth) and the gemma defect (per-branch kill-switch convicted the forced-device residual+RMS arm; drop-the-bake mutation red at op level). The claim file moves to the 2026-08-24 state. The Tenstorrent device-golden pair (our_ids_tenstorrent.npy + neartie_gap_mnats_tenstorrent.npy, teacher-forced via qwen3-neartie-gap-transformers.py on torch 2.7.1+cpu / transformers 5.8.1) describes the FIXED engine's greedy sequence: 247/256 steps inside the 500 mnats near-tie band. The pair is committed now so the remaining red is attributable to the six named out-of-band steps (defect 2b, leg-independent, intrinsic) rather than to missing artifacts; the full gate stays open until 2b closes. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…to the qkvz projection output (mudler#1715) Adds two env-gated debug hooks to the Qwen3.5 paged forward: DumpGdnStage inside GdnBlockPaged (mixed/conv/postconv/core/gated intermediates under $VT_DUMP_ACT, numbered per invocation) and a pre-layer dual-stream snapshot beside the existing per-layer residual dump. The hooks localize defect 2b on prompt 1: pre-layer inputs are bit-identical between the CPU and Tenstorrent arms; post_input_norm is exact (max 0.0098, corr 1.0 — the gemma bake holds); the qkvz projection output already carries scattered gross errors (max_abs 2.1, sign flips, ~0.3% of elements), reproduced identically under VT_GDN_IN_BF16=0. The bf16 activation path and dtype policy are exonerated; the suspect is the projection GEMM on TT at [T,1024]x[1024,6144] with real layer-0 weights. Next: op-level replay of that shape against the CPU result, then read the winning kernel's accumulation path. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
… capture-reliability finding (mudler#1715) Adds VT_DUMP_QKVZ per-invocation captures (mixer input h, resident merged qkvz weight, BA resident weight, and per-call x/out at the MatmulBTRawD boundary) and records what they proved and disproved. Proven by isolated replay with real captured bytes: kMatmulBT is clean at the engine's exact shapes and dtype combinations, including the split-arm BA signature with the REAL h0/w_ba bytes — worst errors are bf16-envelope. Disproven by behavior: the stage-dump claims of garbage projection/BA outputs cannot describe the live dataflow — the engine stays fluent where the captures read 1e38/NaN, and VT_POOL_BYPASS=1 output is identical to the pooled run, so pool-block state is not the carrier either. The capture path itself (DBuf-tmp download under host-free decode, likely via the d2d clone interaction) can serve bytes the live chain never consumes. Every named-kernel claim derived from those dumps is withdrawn; the behavioral ground truth stands (247/256 in-band, six out-of-band steps, solo == batch, f32 leg reproduces, bypass neutral). Next: trusted-capture re-localization (dual-read verification or =0-leg dumps) before naming any kernel for defect 2b. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
Collaborator
|
The Tenstorrent Qwen3.5 implementation is mergeable and the functional/record gates are otherwise green, but the repository's |
… the mixer-input identity anomaly (mudler#1715) Re-runs the stage localization entirely on the VT_TT_HOST_FREE_DECODE=0 leg where no host-free capture hooks are active. Confirmed under the trusted path: the residual stream diverges from layer 0; layer 0's post_input_norm is BIT-IDENTICAL between arms (both run the host gemma arm there); post_attn_norm diverges (max 3.28, corr 0.965) — the defect is born inside the layer-0 GDN mixer, whose real-pass stages diverge progressively (mixed 4.21/0.82 down to gated 3.98/0.66). Adds the triple-read probe and records a precisely-bounded identity anomaly: the buffer captured at ProjectGdnQkvz differs from the bit-identical post_input_norm by max 4.06 / corr 0.85 SYMMETRICALLY on both arms, at the SAME data pointer, while pre- and post-mixer reads return the original bytes — with slot trace showing one registration and no writers, pool bypass neutral, and isolated replays clean. Both explanatory horns contradict surviving facts; the next session breaks the tie with an in-process three-point hash plus immediate op-replay of the pre-GEMM bytes (no cross-process assumptions). FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…consumed another slice's staging (mudler#1715) The staging cache keyed hits on (base slot, dims): an interior slice view of a packed weight resolved to the base slot and, when a previous view with equal dims had staged, consumed THAT slice's device bytes. Engine proof on Qwen3.5: the `a` projection of every GDN layer computed with the `b` weight rows — TT-a equaled CPU-b at corr 0.99998 while sharing the b input bit-for-bit — poisoning g/beta and diverging the mixer from layer 0. Fix: EnsureDevice2D hits and stores require t.data == slot host base; interior views upload from their own bytes as unregistered transients, and an interior view of a device-current base refuses loudly instead of serving stale bytes. Same base-pointer guard applied to EnsureAffine1D. Permanent regression case runs the engine's exact b-then-a sliced-view sequence twice against numerically far-apart halves; the mutation that neuters the base-pointer check goes red at worst 49.5. Post-fix trusted localization (=0 leg, whole-allocation dumps only): layer-0 divergence shrinks but persists, now pinned on kCausalConv1dFwd (conv output max 1.99 corr 0.999 between arms whose inputs match; the arms take different conv sub-paths — indexed gather vs manual). Next: whole-tensor replay of the conv with real captured inputs. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…weight staging verified bit-perfect (mudler#1715) Adds an ops-layer debug seam (declared in tenstorrent_device.h, defined in tenstorrent_ops.cpp) that stages a tensor through EnsureDevice2D exactly as a consuming kernel would and reads the DEVICE copy back as f32. Model-side VT_DUMP_QKVZ capture now emits both the host master (w_host.bin) and this device readback (w_device.bin) for the merged GDN qkvz resident. First result: TT's device-staged in_proj_qkvz is BIT-PERFECT against the host master across all 8,388,608 elements — weight staging is exonerated, and host masters are byte-identical across arms, exonerating the loader. The spec records a MEASUREMENT RESET: several cross-run divergence numbers from ad-hoc captures mixed dtypes (bf16 bytes parsed as f32) and paired matmul calls across arms that take different projection arms, so those specific numbers are withdrawn. What stands: the behavioral signature, the layer-0 mixer origin on whole-allocation =0 dumps, and the three fixed defects. Next session builds one dtype-explicit dual-read dump utility and redoes the stage localization through it. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…harness (mudler#1715) Adds TrustDump (ops-layer, declared beside DebugDeviceReadbackF32): whole- allocation dumps with a typed header (magic/dtype/rank/dims/numel/verified) and DUAL-READ verification — two independent Synchronize+Copy passes must agree byte-for-byte or the payload is omitted. Wired to the GDN mixer stages (packed/mixed/z/conv/pc_q/pc_v/core/gated) on both arms. First trusted results on the =0 leg: post_input_norm bit-identical; conv matches (max 0.035); gated matches within envelope (max 0.039); pc_q is UNCORRELATED (corr +0.028) yet correctly l2-normalized per head on both arms — and mutual-nearest matching rules out a layout permutation. Since gated sits downstream of ql2 and still matches, one of the two readings is lying about buffer identity; the pool itself is exonerated by a 200k-cycle Get/Put hammer (zero collisions). Next session instruments inside kGdnPostConvKernel at the commit site and captures g/beta, which were never dumped. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…he TrustDump commit; preflight pipe swallowed the red) (mudler#1715) The TrustDump commit was pushed while check-env-doc was red because the gate ran through a pipeline whose exit status came from tail. Own it, repair it, and note the lesson: never interpose a filter between a gate and its consumer when the gate's status is the chain condition. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…r misalignment; mixer chain exonerated (mudler#1715) Adds kernel-commit-site TrustDump captures inside kGdnPostConvKernel (q/k/v/g/beta at the CommitDeviceLogical2D sites, plus the g-chain intermediates) and split/merged BA-projection device-truth captures in ProjectGdnBA, all under VT_DUMP_TRUST. The trusted stage matrix then collapsed: emission-aligned comparison across all 18 layers shows pc_q corr >= 0.9991, gated >= 0.9983 and conv >= 0.9996 on every layer. The earlier "pc_q uncorrelated while gated matches" contradiction was an analysis artifact — sorted()[0] paired CPU layer-0 against TT layer-11 because the arms emit different dump-site mixes. Supporting checks along the way: kernel q_out == model-side dump bit-exact; split-path BA gate inputs read real values on the =0 leg with device==host. Two instrument caveats are recorded in the spec as owed work: TrustDump must refresh host bytes via EnsureHostBytes before its dual read (the ambient leg reads stale copies that dual-read verification cannot catch), and VT_GLUE_FUSE=0 is not runnable on Tenstorrent because GdnConvSplit has no native kernel. With the mixer chain exonerated, the remaining gap narrows to the decode-phase packed path, state carry, or the ambient staleness family itself. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
… under host-free decode (mudler#1715) Root cause of the ambient-leg corruption. Under host-free decode a producer commits device-only (host_current=false); EnsureDevice2D then built its upload buffer from the slot's HOST bytes without refreshing them, uploaded pool-fresh zeros to the device, and marked the slot host_current=true — corrupting both the value and the residency record for every later reader. Behavioral split, same build: ambient emitted "!!!(1, 2, 3, 4, 5" for "The capital of France is" while VT_TT_HOST_FREE_DECODE=0 emitted "Paris.". Emission-aligned trusted dumps localized the zeroing between the embedding (real on both legs) and layer 0's mixer input (all-zero on ambient only). With the one-line EnsureHostBytes refresh before the read loop, ambient without any instrumentation completes coherently ("Water boils at" -> "100°C. If a 100"); the =0 leg is unchanged. TrustDump gets the same refresh: its Copy resolves t.data through whichever memory the address maps to, so ambient dual-read verification could not see stale host bytes. Interim dump instrumentation masked this defect because that same refresh heals exactly the state the defect corrupts; recorded in the spec as a Heisenbug note. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…ncy fix — sacred gate 16/16 on ambient (mudler#1715) Quantified parity against the pinned ROCm oracle greedy_ids on the DEFAULT ambient configuration: exact token cells 135/256 -> 213/256, fully-exact prompts 5/16 -> 10/16. The fixed engine's tokens match the pinned oracle exactly on cells where the stale anchor diverged (prompt[1] tok0 our == oracle == 303 vs anchor 948; prompt[12] tok0 9565 both) — the stale pair encoded the corrupt-ambient zeros, which is the justified-drift proof the gate's own message requires before re-derivation. Re-derived via the sanctioned procedure: VT_DUMP_IDS=1 bootstrap dump of the fresh sequence, then qwen3-neartie-gap-transformers.py teacher-forcing against the transformers secondary oracle. Max teacher-forced gap 0.375 nats; every divergence inside the 0.5-nat near-tie band. Full gate verdict on the board: 16/16 prompts PASS (10 strict token-exact, 6 near-tie-band only), 0 forward-divergent, doctest 146/146 SUCCESS. Mutation proof stands from the previous commit's scratch build: removing the EnsureDevice2D refresh regresses ambient to the exact pre-fix garbage; restored byte-for-byte afterward. Owed, recorded in the spec: the test binary segfaults during teardown after printing its verdict (ttnn deallocate / GraphTracker device-destruction order); verdict unaffected. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
…T-QWEN35 ahead of review Union-resolved the single conflict in scripts/check-gate-commands.py (both sides appended their row ID to the runnable population). No source conflicts; the residency fix, TrustDump refresh, and re-derived TT golden pair carry over unchanged. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
Collaborator
|
Triaging the reds on this draft so you know which are yours.
This is AGENTS.md
Leaving the rest to finish. Ping me when you take it out of draft and I will review the Tenstorrent op delta and the e2e gate properly. |
… break and a false-necessity claim (mudler#1715) Fresh-review findings on 1d07d49, repaired: F1 (medium): TrustDump and DebugDeviceReadbackF32 are defined only in the tt-gated ops TU but model TUs call them under a runtime env check, so a default (non-tt-metal) configure of examples/vllm-cli failed to link with undefined references. Added inline no-op stubs under !VLLM_CPP_TENSTORRENT in tenstorrent_device.h, mirroring the WarmPagedKvShadow pattern. Verified: default OFF-build now links clean (503/503) and the tt-on tree still builds. F2 (low): TrustDump's entry EnsureHostBytes was claimed necessary; it is redundant today because Backend::Copy already refreshes the source (tenstorrent_backend.cpp:66-68, since 7faa9c6). Reworded the comment as defense-in-depth and corrected the spec's instrument-caveat note: the ambient all-zero readings were poisoned-slot truth from the EnsureDevice2D defect plus bf16/f32 decode mistakes in analysis scripts, not an unsound Copy path. The necessity claim is withdrawn; only EnsureDevice2D prevents the corruption class. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:x-preview-f-free [opencode]
lu-zero
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
Qwen3.5-0.8B on Tenstorrent produced corrupt text on the DEFAULT configuration
(host-free decode,
VT_TT_HOST_FREE_DECODEunset): "The capital of France is"completed as
!!!(1, 2, 3, 4, 5whileVT_TT_HOST_FREE_DECODE=0completedParis.. This row lands the root-cause fix, the measurement harness thatfound it, and the re-derived correctness goldens.
EnsureDevice2D(src/vt/tenstorrent/tenstorrent_ops.cpp) built its uploadbuffer from HOST bytes without checking slot residency. Under host-free
decode a producer commits device-only (
host_current=false), so the nextconsumer staged pool-fresh zeros onto the device and marked the slot
host_current=true, corrupting both value and record for every laterreader. Fix: one
EnsureHostBytes(t.data)before the read loop.TrustDumpgains the same refresh as defense-in-depth (review finding F2:redundant today since
Backend::Copyalready refreshes its source).!VLLM_CPP_TENSTORRENTforTrustDump/DebugDeviceReadbackF32: model TUs call them under a runtime env check,and default (non-tt-metal) configure of examples previously failed to link.
tests/parity/goldens/qwen35_greedy_0_8b/{our_ids_tenstorrent.npy, neartie_gap_mnats_tenstorrent.npy}via the sanctioned procedure(VT_DUMP_IDS=1 bootstrap, then qwen3-neartie-gap-transformers.py
teacher-forcing against the transformers secondary oracle).
.agents/specs/tenstorrent-qwen35.mdcarries the full evidence chain,including two methodology post-mortems (a layer-misalignment analysis bug
and a Heisenbug note: interim dump instrumentation masked this defect
because EnsureHostBytes' side effect heals exactly the corrupted state).
Why is the change needed
Issue #1715 tracks the Tenstorrent bring-up of Qwen3.5-0.8B. With this fix the
default configuration generates coherent, factual text where it previously
emitted degenerate output, and the row's sacred near-tie correctness gate
passes 16/16 on ambient for the first time.
How can a reviewer verify it
Behavioral split, same build, same prompt (P150 board, flock'd):
!!!(1, 2, 3, 4, 5100°C. If a 100("Water boils at")VT_TT_HOST_FREE_DECODE=0Paris.Paris.(unchanged)Mutation proof (fresh reviewer, scratch clone): disabling only the
EnsureDevice2D refresh regresses ambient to the exact pre-fix garbage class;
restored byte-for-byte restores green.
Parity vs the pinned ROCm oracle greedy_ids (16 prompts x 16 tokens):
exact cells 135/256 -> 213/256; fully-exact prompts 5/16 -> 10/16; max
teacher-forced gap 0.375 nats (band 0.5); full gate 16/16 prompts PASS
(10 strict-exact, 6 near-tie), doctest 146/146 SUCCESS, 0 forward-divergent.
Link proof: default configure (
cmake -B build -G Ninja) now linksexamples/vllm-cli clean (503/503) where the pre-repair tree failed with
undefined
vt::tenstorrent::TrustDump/DebugDeviceReadbackF32.Fresh review: static review + board mutation + scoped re-review of the repair
both returned PASS with findings NONE after remediation.
What remains unverified or out of scope
(ttnn deallocate / GraphTracker device-destruction order). Verdict
unaffected; recorded as owed debt in the spec, fix to follow separately.
this change does not claim them.
scripts/agent-preflight.shon this host skipstest_ltx25_render_compare(python3-numpy unavailable to system python); CI installs numpy and runs it.
arbitrate the examples-on/default-backend link lane this repair addresses.
Refs #1715
FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:x-preview-f-free [opencode]