fix: make a static pod's tap set visible, and refuse a bad steer with a 400

[hijohnnylin]

Follow-up to #227. Two changes, both about the same thing: which sites a pod
declared decides whether a request is served, and that was neither visible at
startup nor reported cleanly when it went wrong.

1. The declared tap set is in both startup banners

Config banner, before the SAEs load:

Initialized Config with:
  model_id: meta-llama/Llama-3.3-70B-Instruct
  ...
  static_points_summary: sae (resolved once the SAEs load) + resid_post.40

Ready banner, read off the loaded model after resolution:

==== LOADING COMPLETE - SERVING ON http://localhost:5002 ====
  ...
  static taps: read resid_post.0-79 | write resid_post.0-79
  startup took 6m 12s

Two lines because they answer two questions. sae has no address list until the
SAEs load, so the first says that rather than printing an empty set that would
read as "nothing declared"; the second is the truth after the extras merge in. A
pod missing a layer its axes need is now visible at startup rather than on the
first request that needs it.

Ranges because the literal list is unreadable — auto on the 70B pod is 80
addresses per direction. One token per point name with layers grouped inside it,
so a gemmascope set reads mlp_out_post.0-25 resid_mid.0-25.

2. A steer at an undeclared layer is a 400, not a 500 mid-stream

#227 fixed the read side and left the write side asymmetric. An axis reading an
undeclared layer got a 400 naming the layer; a steer writing to one still died
inside generate_steered, well into a StreamingResponse that had already
returned 200:

ValueError: Steered generation asked for resid_post.32, which this engine did not
declare. ... Declared: resid_post.40, resid_post.50

assert_steering_available misses this because it asks whether the pod declared
any write site, which is the generation-only case. What bites is writes
declared, at other layers: a projection cap writes wherever the vector it caps
was fitted, so the pod declared 40 for the readout and was asked to write 32. An
uploaded vector can name any layer, so this is not a set a pod can enumerate.

Both checks now match sites the way the engine does, through the
resid_pre[L] == resid_post[L-1] alias, so a pod that declared one spelling is
not refused for the other — the read check in #227 compared literal names and was
stricter than the engine. They lean permissive on anything unrecognized, since
that degrades to the engine's own refusal (the failure they replace) where a
false refusal would break a request the pod could serve.

Deploy note

The pod that prompted #227 no longer uses STATIC_POINTS_EXTRA: capping writes at
layers no list can enumerate, so it runs STATIC_POINTS=auto — priced at
util 0.90 with the prefill chunk pinned to 2048, which leaves KV at ~47,000
tokens. That lives in gitignored local_scripts/pods.yaml, so it is not in this
diff. STATIC_POINTS_EXTRA stays as the lever for a pod that cannot afford auto
and whose axis layers are a closed set; it now has no user.

Test plan

• ruff check + ruff format --check
• Ranges: empty, single, full run, run plus outlier, grouping by point name,
  repeats, and a layerless address
• Config summary for every STATIC_POINTS shape, including a regression test
  that a named mode is not iterated character by character
• Write check: refusal names the layer and STATIC_POINTS=auto, read and
  write sets asked separately, alias match, and the hooked-pod gate
• Full inference unit suite (555) + pyright, against a byte-identical tree in
  a fully-installed venv — this checkout's apps/inference/.venv is partial
  and cannot load conftest.py
• Not exercised on a real static pod. Both refusals and both banner lines
  are only proven by relaunching the 70B pod on this build

Made with Cursor