fix(dashboard): report the model size the installer recorded

[Hoang130203]

Summary

helpers.get_model_info() guesses the model size from the model name, with
a ladder of \b<N>b\b patterns and a 15.0 GB default when nothing matches.
Checked against every model the tier maps actually select, and their declared
LLM_MODEL_SIZE_MB:

tier-map LLM_MODEL         declared GB  guessed GB  delta
  gemma-4-26b-a4b-it              17.6        18.0  +0.4
  gemma-4-31b-it                  19.3        19.8  +0.5
  gemma-4-e2b-it                   2.7         2.8  +0.1
  gemma-4-e4b-it                   5.2         5.3  +0.1
  qwen3-30b-a3b                   18.2        18.6  +0.4
  qwen3-coder-next                47.4        15.0  -32.4  <-- WRONG
  qwen3.5-2b                       1.2         1.5  +0.3   <-- WRONG
  qwen3.5-4b                       2.8         2.8  +0.0
  qwen3.5-9b                       5.6         5.8  +0.2
  qwen3.6-35b-a3b                 20.6        15.0   -5.6  <-- WRONG

The two large misses are MoE models — qwen3-coder-next and
qwen3.6-35b-a3b — whose names do not spell a parameter count the ladder
recognises, so they take the 15.0 GB default. NV_ULTRA's flagship model is
displayed at less than a third of its real size, and SH_LARGE's is off by
5.6 GB. On a dashboard whose job includes telling you whether a model fits your
VRAM, a 32 GB understatement is the wrong direction to be wrong in.

The number was already there

LLM_MODEL_SIZE_MB is not something I am introducing:

• every tier in installers/lib/tier-map.sh sets it (48500, 21110, 1221…);
• .env.example and .env.schema.json document it;
• settings.py lists it among the tracked keys;
• test_model_activate.py already asserts activation persists it ("21110").

get_model_info() was the one place that did not read it.

Fix

Prefer the declared value, keep the name ladder as the fallback:

• A missing key → ladder, exactly as before, so installs predating the key are
  unaffected.
• A non-numeric (auto, a trailing comment) or non-positive value (0 on the
  CLOUD tier, -1) → ladder, not 0.0 GB.
• Quantization detection moves into a small _detect_quantization() helper so
  both paths share it rather than the new path duplicating the awq/gptq/gguf
  checks.

Verified end to end against real .env shapes:

  coder-next (declared 48500 MB)         size_gb=47.4    quant=GGUF ctx=131072
  35b-a3b (declared 21110 MB)            size_gb=20.6    quant=GGUF ctx=131072
  2b (declared 1221 MB)                  size_gb=1.2     quant=GGUF ctx=8192
  coder-next, no declared size           size_gb=15.0    quant=GGUF ctx=131072
  9b, no declared size (ladder)          size_gb=5.8     quant=GGUF ctx=32768
  malformed declared size                size_gb=5.8     quant=GGUF ctx=32768
  declared size 0 (cloud tier)           size_gb=5.8     quant=GGUF ctx=32768

Test

Adds 10 assertions to the existing TestGetModelInfo: the declared path, the
declared value beating a name the ladder would have matched, both MoE names
the ladder cannot read, the no-key fallback, every unusable declared value
(parametrised over "", auto, notanumber, 0, -1), and that
quantization detection still works on the declared path.

AI Assistance

AI assisted with drafting the fallback logic, the test matrix, and this
description. I built the comparison table above by running the shipped ladder
against the tier maps' own declared sizes, which is how the misses surfaced.

Release Lane

• Stable hotfix targeting release/2.6.x
• Mainline change targeting main
• Next-minor work targeting the next feature/minor release
• Not sure; reviewer should help classify

Stable hotfix reason:

n/a

Changed Surface

• Docs only
• Tests only
• Dashboard UI
• Dashboard API / host agent
• Installer / bootstrap / lifecycle
• Docker Compose / service manifests
• Model routing / Hermes / capabilities
• Network exposure / auth / proxy
• Dependencies / runtime wiring

(One function in helpers.py plus a small extracted helper, and its tests.)

Risk And Validation

• Risk level: Low
• Validation run:
  • git diff --check
  • Markdown/link sanity for docs
  • Focused tests listed below
  • Dashboard lint/test/build
  • Extension audit / compose validation
  • Release-grade fleet or scoped hardware validation
  • Stable-lane patch validation, if targeting release/2.6.x

Commands/results:

$ python3 -m py_compile extensions/services/dashboard-api/helpers.py
py OK

$ python3 -m pytest tests/test_helpers.py -q
116 passed

$ python3 -m pytest tests/test_helpers.py -q -k GetModelInfo
24 passed, 92 deselected

# the new assertions against origin/main's helpers.py:
FAILED TestGetModelInfo::test_prefers_the_declared_size_over_the_name_ladder
FAILED TestGetModelInfo::test_declared_size_wins_even_when_the_ladder_would_match
FAILED TestGetModelInfo::test_moe_model_the_ladder_cannot_read
FAILED TestGetModelInfo::test_declared_size_keeps_quantization_detection
4 failed, 20 passed

No pre-existing test changed; the 14 that were already in TestGetModelInfo
still pass unmodified, which is what pins the fallback behaviour.

Operational Change Check

get_model_info() is a read path behind the dashboard's system-status view. The
change alters one reported number, and only towards the value the installer
recorded. Nothing is written, no schema changes, and an install without
LLM_MODEL_SIZE_MB in its .env behaves exactly as before — which is what the
untouched existing tests assert.

• This is not an operational change.
• This is an operational change and validation is recorded above.
• This is an operational change and validation is intentionally deferred for:

Notes For Reviewers

The ladder is now the fallback, and it is still wrong for those two models.
I deliberately did not add \b35b\b and a coder-next special case: patching
the ladder treats the symptom, and any model added tomorrow reintroduces it.
The declared value is the fix; the ladder only has to cover installs that
predate the key. If you would rather delete the ladder entirely and return
None when the size is unknown — arguably more honest than a 15.0 GB guess —
that is a smaller diff than this one and I am happy to send it instead.

Adjacent, not fixed here: the same .env read also parses CTX_SIZE /
MAX_CONTEXT with a 32768 default. That default is not obviously right for
Tier 0 (8192) or NV_ULTRA (131072), but it only applies when both keys are
missing or unparseable, which the installer does not produce. I left it alone
rather than widen the diff.