stack 7/7: NIM vision classification, service repair, and the qwen3.8-max rename - #980

The A-gate reviewer returned FAIL. Every blocker was reproduced before being
accepted; none was rebutted. 001 records the synthesis.

B1 killed my own design. I proposed maintaining the 15 vision-capable ids and
deriving text-only as their complement, and claimed an unclassified model would
default to sidecar-on. It does not — a complement over a static chat-model list
leaves an unknown id in neither list, so modelInList returns false and #956
survives verbatim:

  deepseek-ai/deepseek-v4-flash       sidecarWouldRun=true
  moonshotai/kimi-k2.6                sidecarWouldRun=false
  brandnew/model-nobody-classified    sidecarWouldRun=false

I had inverted which list is maintained while keeping the closed world — the
same lesson as the three earlier allowlist failures, reproduced while writing
the document that cites them. 010 now changes the predicate instead: default-on
for the nvidia entry with the vision list as its exception set, so a stale
exception list costs one description hop rather than reproducing the bug.

B2: removing a native-vision id from noVisionModels is not sufficient. The
catalog advertises image input only for list members, so those models would be
blocked client-side instead. They need explicit modelInputModalities.

B3: src/update/job.ts:775 skips the service refresh entirely on non-elevated
Windows — the dashboard path. Its stated reason is that schtasks /create needs
UAC, which repair does not call, so the skip must be narrowed or the reporter's
own surface stays broken.

B4: repairService throws plain Errors and bin/ocx.mjs sees only an exit status,
so 'fall back on not-installed' was unimplementable. Re-run diagnoseService()
after a failed repair instead of parsing messages.

B5: retargeting emits 'edited', which ci.yml does not listen for, so a green
check on the same head sha proves nothing about the new merge base.

030 also moves the #964/#970 closure from 'when stack 7 opens' to 'open and
green'. The earlier text borrowed a policy from the six carried PRs, which had
verified replacement commits already on a branch; this replacement does not
exist yet and its first design just failed audit.

…ler honest design

Audit round 2 closed B2/B3/B5 and returned FAIL on two P0s. Two consecutive
failures on the same surface means root cause, not a third patch of the same
shape.

R2-B2 is the one that matters: 010 flagged 'non-chat endpoints never reach the
predicate' as a thing to confirm rather than assume, and I did not confirm it.
It is false. NVIDIA has no discovery filter and shouldExposeRoutedModel rejects
only media-generation names, so embeddings, rerankers, guards and OCR all reach
planVisionSidecar:

  nvidia/nv-embedqa-e5-v5                        filteredOut=false
  nvidia/llama-3.1-nemotron-safety-guard-8b-v3   filteredOut=false
  nvidia/nemotron-ocr-v2                         filteredOut=false

Under default-on every one of them would advertise image input and burn a
sidecar call before failing upstream.

Root cause: twelve of the thirteen registry entries declaring noVisionModels
pair it with a static models list. NVIDIA is the first asked to classify over an
unbounded set, with no modality and no model-kind metadata. An unknown NIM id
therefore carries no signal separating a text-only chat model from an embedding
endpoint, and no predicate over an id string can recover information the
provider does not publish. Draft 1 kept the closed world; draft 2 escaped it but
claimed knowledge that does not exist.

The design that follows: enumerate the known text-only ids (correcting #964's
five false positives), pin the 15 verified vision ids with explicit
modelInputModalities so they become usable, leave unknown ids untouched, and
record the open-world gap as a stated limitation. Confined to registry.ts with
no predicate change, so no consumer edits — the reason this draft is
implementable where draft 2 was not.

R2-B1 also caught two consumers earlier drafts missed: web-search/index.ts:165,
and cli/models.ts:44 which uses raw .includes() instead of modelInList.

R2-B3 (bin/ocx.mjs cannot import diagnoseService from TypeScript) was found and
fixed before the verdict arrived; 020 already reads startup.serviceInstalled
from the status --json subprocess it spawns.

Audit round 3 FAIL. Three failures on one document is LOOP-DOOM territory, so
this changes the verification method the design rests on rather than patching
the design again.

R3-B1: moonshotai/kimi-k2.5 is a sixth false positive in #964's list — NVIDIA
documents GIF/JPG/PNG input, four images per prompt, with hosted image_url
examples. This is fatal to draft 3's justification, not just a missing entry.
Draft 3 argued 'for a known id the classification is real and verifiable' while
inheriting ~54 unaudited entries from #964 and calling them known. Finding a
sixth immediately after correcting five proves I never verified the remainder.
Every carried id now gets verified against NVIDIA docs or dropped; dropping
costs today's behavior, assuming costs a silent regression.

R3-B2: my registry census was wrong. Counted directly there are 17 entries
declaring noVisionModels, not 13, and the two without a static models list are
opencode-go and opencode-free — opencode-zen declares none at all. The numbers
came from an ad-hoc regex whose entry boundaries were wrong, and I wrote its
output into two documents as fact. Same failure as R2-B2, one document later.
The information-constraint argument survives and the two real exceptions
strengthen it: both classify only known ids, and opencode-free has a -free
suffix filter NVIDIA lacks.

R3-B3: test 5 asserted that a user's noVisionModels 'wins' over the registry.
mergeStringArray unions them, so a user cannot remove a registry entry. Test
now asserts additions are preserved.

R3-B4: dropped the dated snapshot test. A local date assertion has no NVIDIA
input, so it detects elapsed time rather than drift, and its cheapest CI fix is
bumping the date without auditing anything.

Also: 030 now requires #956 to close with an explicit bounded-scope statement,
and 020 records that the status probe runs only on the success path today.

… dropped

003 made per-id verification a gating step. This is that audit, run against
build.nvidia.com model pages and the NIM LLM/Visual API indexes on 2026-08-04.

#964 submitted ~64 ids. Fewer than half survive:

  26  confirmed text-only (explicit 'Input Modalities: Text') — these ship
   6  confirmed image-capable — moved to the vision list
  32  unverified or absent from NVIDIA's catalog — dropped

The 26 include z-ai/glm-5.2, deepseek-v4-flash/pro and the nemotron-3 family,
so the models issue #956 actually names are all fixed.

No seventh false positive was found, which is the first evidence the correction
has converged rather than merely advanced.

The 32 dropped are mostly delisted models — harmless in isolation, since nobody
can route to a model NVIDIA no longer serves. But the set includes
nvidia/nemotron-nano-3-30b-a3b, a reversed-name typo of the real
nvidia/nemotron-3-nano-30b-a3b which the same list also spells correctly, and
mistralai/mixtral-8x22b-v0.1 where NVIDIA documents mixtral-8x22b-instruct-v0.1.
Half the list was assembled rather than verified; the six reversed entries were
the visible damage, this is the extent of it.

Kimi is now split correctly across two independent axes: k2.5 and k2.6 join the
vision list, k2-thinking and k2-instruct stay text-only, and all four remain in
NVIDIA_NIM_KIMI_MODELS for reasoning suppression.

google/codegemma-7b verifies while google/codegemma-1.1-7b does not — adjacent
names, opposite outcomes, which is why name-based classification was rejected.

…uals

Adds 040: qwen3.8-max-preview becomes qwen3.8-max, and the price overlay moves
from a Routeway reseller proxy to Alibaba's published rate.

Alibaba released Qwen3.8-Max as stable on 2026-08-03 and documents the preview
endpoint as liable to be taken offline. Model Studio lists both ids today, so the
rename touches 10 sites in registry.ts, 3 in expected-prices.ts, and 6 test
files across both alibaba-token-plan providers.

No -preview alias is added. A config naming the old id still routes, because
routeModel accepts an arbitrary namespaced id for a configured provider and the
upstream still serves it; what such a user loses is capability metadata keyed to
a retiring preview id, which is the correct outcome.

Price: Qwen publishes $2 input / $6 output. Two honesty constraints recorded —
the figure is Qwen's own announcement and Model Studio has no qwen3.8-max row
yet, and cache rates are unpublished so both cache fields go to 0 rather than
inheriting the Routeway numbers. Carrying a reseller cache rate under a vendor
price label would be a wrong number wearing a verified badge. The Routeway
constant and its overlay are removed entirely, which is exactly the exit
condition its own comment named.

Round-4 NEAR-PASS residuals closed:
- k2.5 raised the vision set to 16 and the reversed-entry count to six; both
  numbers were stale in 000, 010. The k2.5 regression test was missing from the
  no-sidecar and emitted-modality cases and is now required in both.
- 011 claimed a delisted id 'cannot be routed to'. False: routeModel accepts
  arbitrary namespaced ids and a stale cache can surface one. The disposition
  holds for a narrower reason — exclusion leaves them at today's unclassified
  behavior — and the text now says that instead.
- The Mistral Medium 3.5 hosted-endpoint recheck is resolved; it ships.

The nvidia registry entry declared no noVisionModels, so planVisionSidecar never
fired for any NIM model and the catalog never advertised image input. A text-only
NIM model therefore either received raw image parts it cannot read, or had
attachments blocked client-side. That is issue #956.

Two verified lists, both audited per-model against NVIDIA documentation on
2026-08-04 (evidence: devlog/_plan/260804_stack7_service_vision/011_nim_id_audit.md):

  noVisionModels          26 ids — text-only, sidecar describes their images
  modelInputModalities    16 ids — natively image-capable, native path, image
                                   input advertised explicitly

PR #964 proposed ~64 text-only ids. Six are natively image-capable per NVIDIA's
own docs — inkling, minimax-m3, kimi-k2.6, kimi-k2.5, step-3.7-flash and
mistral-medium-3.5-128b — and listing those is a silent defect: the model can
read the image, but the proxy substitutes another model's text description. No
error, worse answers, extra cost. Issue #956's body carries two of the same
errors. A further 32 of #964's ids have no current NVIDIA page and are dropped
rather than assumed text-only.

The vision list also needs explicit modelInputModalities. Removing an id from
noVisionModels is not enough: the catalog advertises image input only for list
members, so a natively-capable model would be published as text-only and the
Codex app would block attachments before the native path could run.

Unclassified ids are left alone deliberately. NIM publishes no modality metadata
and shouldExposeRoutedModel filters only media-generation names, so embeddings,
rerankers, guards and OCR endpoints reach this same path — an unknown id carries
no signal separating them from a text-only chat model. Defaulting in either
direction would be a claim the data does not support.

Vision and reasoning stay independent: k2.5/k2.6 join the vision list while
k2-thinking/k2-instruct stay text-only, and all four keep reasoning suppression.

Red-green: reintroducing #964's kimi-k2.5 entry fails 2 guards; dropping the
modalities map fails 3; removing noVisionModels entirely fails 8. Restored:
23 pass / 0 fail.

…it (#970)

`ocx update` stops the proxy before replacing package files, then brought the
service back with `ocx service install`. The Windows scheduler installer always
reaches `schtasks /create`, which requires elevation the updater does not have,
so an ordinary non-elevated update stopped a working proxy and could not restore
its managed service.

serviceReinstallArgs() now returns ["service", "repair"], which rewrites the
wrapper assets and restarts the EXISTING registration without /create. The
export name is kept for out-of-module callers; serviceInstallArgs() is split out
for the paths that genuinely need to register.

Safety of the substitution: repairService() throws when the service is not
installed, and the update path runs after `ocx stop` — but stop never
deregisters on any platform. macOS unloads the plist, Windows calls /end, Linux
calls systemctl stop; deletion lives only in the uninstall paths. Verified
across all three (evidence: devlog 020).

Two things a straight argv change would have missed:

The Windows GUI worker skipped the refresh entirely (update/job.ts) because its
own comment said /create would UAC-fail. That reason does not survive repair, so
the skip is narrowed to callers still passing install argv — otherwise the
dashboard-triggered update, the most common Windows path, keeps the bug while
the CLI gets fixed.

bin/ocx.mjs infers 'a service manages this proxy' from service-state.json
existing, which can be stale. Repair correctly refuses that case, but its thrown
Error is indistinguishable from any other failure there (plain Error, inherited
stdio, generic exit status), so message-matching was unimplementable and a
blanket install-on-failure would resurrect the elevation prompt. It now reads
startup.serviceInstalled from the `status --json` subprocess it already spawns —
the file is plain Node ESM and cannot import diagnoseService() directly.

Advice strings that fire only for an INSTALLED service now say repair: cli/status,
winsw missing-binary, stale baked paths, stale scheduler assets, the launchd
older-plist and not-loaded hints. First-install and missing-unit guidance stays
install.

Red-green: restoring the unconditional Windows skip fails the new guard. Six
existing tests pinned the install argv and were updated with reasons.

245 pass / 0 fail across the service, update, winsw, doctor, status, startup and
Windows-deploy suites.

…it from the vendor

Alibaba shipped Qwen3.8-Max as a stable model and documents the preview endpoint
as liable to be taken offline once preview concludes. Model Studio lists both ids
today, so this moves the registry to the supported one across both providers:
10 sites in registry.ts, the price overlays, and 11 test files.

No -preview alias is added. A config still naming the old id keeps routing —
routeModel accepts an arbitrary namespaced id for a configured provider and the
upstream still serves it. What such a user loses is capability metadata keyed to
a retiring preview id, which is where that metadata should no longer live.

Pricing moves from a Routeway reseller proxy (1.5/5/0.15) to Qwen's published
$2 input / $6 output, which is exactly the exit condition the old overlay's own
comment named. Two caveats stay in the source string rather than being dropped:

- the figure is Qwen's release announcement, not an Alibaba Model Studio billing
  row (Model Studio still lists qwen3.7-max and qwen3-max, with no 3.8 entry);
- no cache rate is published anywhere, so both cache fields are 0 rather than
  inheriting the reseller's 0.15. A reseller number under a vendor-price label
  would be a wrong value wearing a verified badge.

Status rises to 'verified' for input/output because the vendor published them.

The intl provider's defaultModel stays qwen3.7-max — that predates this change
and renaming an id is not a licence to change which model a provider selects.

Red-green: dropping any single metadata key during the rename fails the new
survival guard. Full suite 7753 pass / 8 skip / 0 fail across 508 files.