[Whq02]

What this changes

Two places in the docs say something the server does not do. Both were read beside the server's own JSON (1.2.2 on a
Mac with Qwen3.5-2B) and the code on main at 342b237.

1. docs/api.md's health table says memory is {} on Metal. A Metal family with a memory reader answers the process
   footprint (the 2B, served with --prompt-cache-gib 0):

   $ curl -s http://127.0.0.1:8080/health
   {"status": "ok", "model": "qwen3.5-2b", ..., "memory": {"active": 2837990136, "cache": 0, "peak": 2837990136}, ..., "backend": "metal", ...}

status_routes.zig writes whatever engine.memory() reads and {} when it reads nothing. cache_fit.hostMemory is
the reader of Nemotron (metal.zig), the 2B (qwen35.zig), the 27B and Qwen3.5-MoE (qwen27_served.zig, the host
metal.zig opens both through), Flash Next (flashnext_batch_host.zig, flashnext_host.zig), GLM-5.3-Flash
(glm_host.zig) and GLM-5.3 (EXLR8) (glm_dsa_host.zig); Flash Next and the two GLMs pass their load limit, which
the route writes as limit. The families without a reader (Gemma 4, LFM2.5, LFM3, Llama, MiniMax, Spark, Step-5,
Kimi, Qwen4, and Drama 3, whose drama_host.zig reads memory on CUDA only) are the ones that answer {}. The row now
says that, keeping the limit clause the table already had.

2. docs/memory-diagnostics.md lists Nemotron, Flash Next and Qwen3.8-27B as the families that report
   prompt_cache_plan. On a Mac the hosts that publish it are those three and Qwen3.5-MoE (qwen27_served.zig, the
   27B's host: the plan when the budget is positive, omitted at 0, as the page already said of the 27B), Gemma 4
   (gemma4.zig: the same rule), LFM2.5-24B-A2B (lfm2_moe.zig) and LFM3-1.7B (lfm3.zig), both at any budget, 0
   included, and the 2B with an explicit --prompt-cache-gib (qwen35.zig: the plan at any budget, the store only when
   it is positive). With --prompt-cache-gib 0:

   $ curl -s http://127.0.0.1:8080/memory
   {..., "prompt_cache_plan": {"source": "physical_footprint", "budget_bytes": 0, "explicit_budget": true, "over_cap": false, "ram_bytes": 51539607552, "ready_footprint_bytes": 3356707864, "cap_bytes": 36077725250, "room_bytes": 30332138026, "margin_bytes": 2147483648}}

The table row and the zero-budget sentences now name them; the sentence #2367 added about the 2B keeping prompt states
only with --prompt-cache-gib stays, with the budget_bytes: 0 case beside it.

Nothing else on either page changes.

Mercury CLI integrates TensorFold and reads /health and /memory to show what a served model holds; it was written
to the sentences and then read the JSON.

• Issue: n/a: a docs correction, with the server's output above.
• Stacked on: main (342b237), two commits: the docs commit and the CHANGELOG line under Unreleased with this number.
• atc overlap --json: run 2026-11-07T12:17Z on the previous head against origin/main at 16e044d; rebased once more onto 342b237 (two Unreleased lines landed meanwhile; this line now sits at the bottom of the block, git merge-tree clean) (ATC 0.9.0; complete: false, pullFilesLimit). changedFiles 3. On the two pages: ten open PRs share docs/api.md (GLM-5.3-Flash on CUDA keeps conversation states on disk, every rank's part (--learn-conversations) #1411, lfm3: add native Metal vision support above the text decoder #1502, Nemotron 3.5 Transcribe: audio input, the 65,536-token window, an hour of audio, MTP drafts #1632, dsv41: serve structured output with the lane core's grammar masks #1665, server: an explicit thinking off that a template can't honour is refused #2170, server: POST /v1/images/generations for an engine with an image capability (#1573) #2253, moss image: serve pictures on CUDA (encoder, decoder and host for #1573) #2281, Qwen3.8-27B on a Mac: the disk staging permit takes one whole state at its context #2388, GLM on CUDA: the latent split's per-rank page book and owner (TF_GLM_KV_SPLIT, #1431 part 1; stacked on #1411) #2415, docs: the end-token pair for replies that quote a template marker #2437); git merge-tree of each against this head and against main alone gives the same result for every one (lfm3: add native Metal vision support above the text decoder #1502 conflicts in docs/api.md and its LFM3 files either way; server: POST /v1/images/generations for an engine with an image capability (#1573) #2253, Qwen3.8-27B on a Mac: the disk staging permit takes one whole state at its context #2388 and docs: the end-token pair for replies that quote a template marker #2437 merge clean; the others conflict in their own files either way), so nothing conflicts on the pages because of this change; no open PR changes docs/memory-diagnostics.md. On CHANGELOG.md: the read lists lfm2_moe: serve LFM2.5-24B-A2B on CUDA (plain serve, shared kernels) #1492, Nemotron Parse on HIP: NVIDIA's checkpoint, token-exact against its HF example #1941, LightOnOCR-3 (0.8B, 1B, 4B) on native HIP, with the model card's oracle #1965, profile: register the pool-page switches for Dory, Gemma 4, Spark2.5 and Nemotron's Metal paging #2306, feat(hip): tap a drafter's layers into each stream's ring #2314, qwen3_5 cuda: readout streams fill in pieces of final-norm rows (#2324 step 2) #2384, CUDA: audit sm_110 cubins without widening release records #2395 and docs: the end-token pair for replies that quote a template marker #2437 as conflicts, each in CHANGELOG.md only (every one adds its own line to the same Unreleased block), the same shape as cli: tensorfold models --json, the listed checkpoints as one document #2364 and server: a refused API key file logs the path and the mode it read #2365 had.

Checks on this head

• Head commit: 443e1a0
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: python3 tools/pr_check.py --base origin/main --pr 2439 on this head: merge base 342b237, 3 changed files, lean 116 on the base and 116 on this head, 0 new; 0 errors, 0 warnings. The pages are outside the checker's trees (zig, tools/zig); no Zig file changes.
• Tests run, with the command and pass, fail and skip counts: git diff --check origin/main..HEAD clean; zig build test-golden --summary all on the docs commit's tree (the two pages unchanged since): Build Summary: 5/5 steps succeeded, 322/335 and 55/55 equal to the server goldens, 0 unexpected differences (nothing in the tree reads these pages; the 13 are its recorded known differences).
• CI: read on this head after the push; the first run on 281721b (the same docs change, before the CHANGELOG line and the identity fix) had benchmark-clients (ubuntu-latest) and zig-macos-dev-tools green when the head moved.
• Not run, and why: zig build test; no source changes.

Receipt

n/a: documentation; the /health and /memory bodies this corrects are quoted above.

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. (n/a: no test; the receipt is the server's JSON beside the sentence.)
• 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.