[jishnuvenugopal]

Thank you, @ashhart, for the named host steps and the recording devices behind them (mtl.Record, the fake CUDA drivers): they
check drafted trees, residency after a failed allocation and chunk grids with no GPU at all, which is what made each of these breaks
easy to pin to one commit. And thank you for the #1327 review that settled GLM's chunk grid; one of these tests just had not caught up.

What this changes

Six host steps fail on main because a later change moved code under a test fixture that was not updated with it. Each commit
updates one fixture to the change that moved it, in tests only, except one line: qwen3_5/cuda.zig now exports cuda_cache.zig
as cache, beside engine and draft, so the shared-stream test can count pendBytes instead of copying its formula.

Commit here	Change on main it follows	What the fixture missed
qwen35 drafter fixtures	3e88d9b: DFlash2's facts read the target's shape (copyMatch(d.e.g))	Draft fixtures never set d.e, an unset pointer
Nemotron residency fixture	d643104: kv_arena, kv_tables, kv_pages, kv_pipes on the backend	Metal = undefined left them undefined
Qwen27 frame replacement	ede6249: a frame slot reserves 64 KiB	a 1 KiB replacement fit the old slot, so nothing was replaced
GLM opt-in test	7bf9711 (#1327): GLM's step takes the cap hint as unused	the test still expected the off-grid BadChunk
qwen35 Keeps fixtures	58c5085: Keeps.apply_pend, and streamBytes adds pendBytes	six Keeps literals and one stream-size sum

The two Draft fixtures get a small target shape (hidden 256, one layer), so copyMatch returns 0, the target's own copy
match, as before 3e88d9b. The GLM test keeps its opt-in setup and its check that the 512-row cap reaches the driver, then asserts
the property your review settled: every chunk start stays on the 2048 grid.

• Issue: Refs CI never reaches 246 of the 393 named test steps, and nine of them fail on main #2031, which proposes running these steps in CI. These are test fixtures behind changes already on main.
• Stacked on: main at 6a50486.
• atc overlap --json: not run (ATC is not installed on this machine). The same check by hand: the open pull requests that touch
  these files are cuda: tune dense 27B MLP tiles on RTX PRO 4500 #1730 (qwen35_draft_frame_host.zig) and qwen3_5 cuda: a kept path longer than pend_cap lands at once instead of overrunning the pending buffer #1982 (qwen3_5/cuda.zig); git merge-tree with each is clean.
  cuda: tune dense 27B MLP tiles on RTX PRO 4500 #1730 also fixes qwen35_live_host.zig's "served scratch" fixture, the same 3e88d9b break, so this branch leaves that one to it.

Verified outcome

• Why: six of the named test steps that check drafting, buffer residency and prompt chunking fail on main: three crash or fail,
  and three no longer compile, so the checks they were written to make have not run since late October.
• What was done: each test's setup was brought up to date with the code change that moved under it; no engine behavior changes.
• Result: all six steps pass (152 tests) where main has 6 crashed or failed tests and 3 builds that stop, and the repaired
  tests again catch a planted break of what they check.

Checks on this head

• Head commit: 6f3c2c8.

• Lean checker on the base (6a50486) and on this head: 424 and 424, 0 new.

• Tests run, uncached (--cache-dir on an empty directory), M1 Pro, macOS 26.6:

  Step	main 6a50486	this branch
  test-qwen35-draft-trees-host	1 pass, 3 crash	4 pass
  test-qwen35-draft-frame-host	does not compile: missing struct field: apply_pend x3	12 pass
  test-qwen35-rounds-host	does not compile: apply_pend x1	8 pass
  test-qwen35-shared-host	does not compile: apply_pend x2	13 pass
  test-metal-residency-owners-host	66 pass, 1 fail, 1 crash	68 pass
  test-glm-scheduling-host	46 pass, 1 fail	47 pass
  test-qwen35-frame-invalidation-host (shares an edited file)	2 pass	2 pass

• Each break pinned to one commit: on its parent the step passes, on the commit it fails (3e88d9b40^ 4/4 and 14/14, 3e88d9b40
  3 and 1 crash; d64310496^ 65/65, d64310496 1 crash; git bisect run from d643104 to main finds ede6249 for the Qwen27
  test). The GLM test already fails at 7bf9711, where it landed, so it is not bisectable.

• Planted breaks, each run against the repaired step: Qwen27's replacement keeping the old buffer in the residency set (caught);
  GLM's step honoring the hint (caught: BadChunk); residency.openCache without errdefer cache.deinit (caught); streamBytes
  without pendBytes (caught). Not caught: Nemotron's residency.apply without &b.scratch; that gap was already there and this
  branch does not change it.

• zig build test --test-timeout 10m, uncached, on main and on this branch: 8336 of 8375 tests pass, 39 skip, 0 fail on both; the
  same 55 steps fail on both, each a Metal library build that needs Xcode on this Mac.

• CI's other steps, on main and here: test-golden 375 equal, 0 differ, 11 known on both; test-kernel-contract 138/138,
  fuzz-cortheon 19/19 and test-server-security 19/19 on both; test-pull-security stops on both at the same 44 Metal library
  builds. Plain zig build: the same 64 steps fail on both, all Metal library builds.

• Sweep of all 393 test-* steps on main, one zig build each: 361 pass, 23 need Xcode's Metal compiler or -Dxgrammar, and 9
  fail in code: the six above, test-qwen35-live-host (cuda: tune dense 27B MLP tiles on RTX PRO 4500 #1730), and test-gradient-serving and test-qwen-serving-host, whose
  reviewed-source hashes no longer match (ck_experts.cu, gdn.cu). Those two are review pins, so this branch leaves them to you.
  The same sweep on this branch differs from main in exactly the six steps above; the other 387 keep their result.

• zig fmt --check build.zig zig and git diff --check origin/main: clean.

• CI: runs test, test-golden, test-kernel-contract, test-server-security, test-pull-security and fuzz-cortheon, none of
  which builds these steps, so CI cannot show this change either way.

• Not run: CUDA devices, and Linux or Windows hosts.

Receipt

n/a: tests and one module export, nothing that reaches inference.

• Environment: main 6a50486, Zig 0.17.0, Apple M1 Pro, macOS 26.6 (Command Line Tools, no Xcode). No checkpoint.
• Paths and platforms: host tests only (zig/build/cuda_family_host.zig, the residency and GLM scheduling steps). No kernel.
• Exactness, decode speed, prompt speed, licence: n/a, no engine change.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new tests: each repaired test
  fails or does not build on main and passes here; the planted breaks above show they test their behavior again.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @jishnuvenugopal, this is a clean, careful fix. I checked it on current main: all five commits still apply
cleanly, the lean count stays at 424, and each of the six steps still fails there for the reason you give. The
copyMatch target, the apply_pend field, the four Nemotron KV fields, the 64 KiB frame slot and GLM's ignored cap
hint are all still on main. Nothing landed in the meantime covers them, so every fix here is still needed. Each test
still asserts what it was written for: the Qwen27 test now reaches the replacement it meant to check, and the GLM test
keeps its cap check while asserting the grid GLM actually keeps. The one production line, exporting cache so the
shared test reads pendBytes, makes sense. Thank you also for pinning each break to its commit and planting the
breaks to show the steps catch them again. We'd like to land this soon, and #2031 would keep these steps from going
stale again.

[ashhart]

Thank you @jishnuvenugopal. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @jishnuvenugopal! This is in TensorFold 1.2.3, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.3