fix: let a static pod declare the layer a readout axis reads

[hijohnnylin]

What broke

The 70B pod 500s on any assistant-axis request:

ValueError: Capture during generation asked for resid_post.40, which this engine
did not declare. It was built with backend='vllm-static', whose tap set is fixed
when the CUDA graphs are recorded, so no new tap can be installed on the running
model. Declared: resid_post.50

STATIC_POINTS=sae declares exactly the sites the loaded SAEs read, and
resid-post-gf is a single SAE at layer 50. The lu_assistant-axis row for
llama3.3-70b-it reads layer 40.

This only became possible in #226. A shipped asset was on the pod's disk before
the graphs were recorded, so which layers it would be asked for was a deploy-time
fact; an axis that arrives with the request makes it a per-request one, and
nothing on the pod can derive it at startup.

The 8B pod is unaffected — it runs auto, which already covers its axes at
layers 13 through 29.

Why not an existing value

Neither alternative fits one extra layer:

• sae+auto covers it, but declares every layer. pods.yaml records this exact
  pod as "could take the flip and is refused it on memory"; one layer costs
  1/80th of it.
• An explicit JSON list is documented as the only value that declares no writes,
  so it would trade every steer for the readout.

The change

STATIC_POINTS_EXTRA adds sites beside a set the server resolves for itself
(sae / sae+auto), as a JSON list in the same spelling — ["resid_post.40"].
Declared as reads and writes, for the reason auto implies its writes: a
capture site that cannot be written is a readout that works and then refuses
every steer at the same layer, and steering on a persona direction at the layer
it was fitted at is a thing this server is asked to do. Deduplicated, so naming a
site an SAE already covers is free. Refused rather than ignored on any other
STATIC_POINTS value, where it would be a no-op that reads, in a deploy config,
as though a site had been added.

The refusal moves earlier and gets specific. assert_residual_available asks
whether a point is declared, which is right for an endpoint that reads every
layer or none — an axis reads exactly one, so resid_post looked present and
only the address was missing. assert_capture_layers_declared checks by address
before anything is rendered or generated, and returns a 400 naming the axis, the
missing layer and the variable to set, instead of a traceback out of the engine.

The pod's own STATIC_POINTS_EXTRA: '["resid_post.40"]' lives in gitignored
local_scripts/pods.yaml, alongside a sae_memory.py change so the freeze-buffer
estimate counts the two extra buffers (32 KiB/token on d_model 8192).

Test plan

• ruff check + ruff format --check
• Unit tests for the parse (both address spellings, malformed input, unset),
  the union (reads and writes, dedup), and the 400 (by layer, not by point
  name; a hooked pod still captures anywhere)
• Full inference unit suite + pyright, run against a byte-identical tree in a
  fully-installed venv — this checkout's apps/inference/.venv is a partial
  install with no torch, so pytest cannot load conftest.py here
• Not yet exercised on a real static pod. The layer-40 capture is only
  proven by relaunching inference-llama-3.3-70b-it-static on this build

Made with Cursor