feat(aimlapi): add aimlapi.com provider

[Lookoff-AIMLAPI]

What

Adds aimlapi.com as a provider to the catwalk database, so Crush users on
AI/ML API get a maintained model list instead of hand-writing a custom
openai-compat block.

Two commits, deliberately separable:

Commit	Contents	Upstream-safe?
feat(aimlapi): add aimlapi.com provider	the provider itself	yes
chore(aimlapi): fork-only placement — do not send upstream	moves us to the head of providerRegistry	no — drop this commit before offering anything upstream

The provider commit

• internal/providers/configs/aimlapi.json — type: openai-compat,
  api_endpoint: https://api.aimlapi.com/v1, api_key: $AIMLAPI_API_KEY,
  9 hand-curated chat models.
• internal/providers/providers.go — //go:embed, loader func and registry
  entry, in the alphabetical slot after aihubmix.
• pkg/catwalk/provider.go — InferenceProviderAIMLAPI const and its
  KnownProviders() entry. (atlascloud provider: add atlas cloud charmbracelet/catwalk#463 and neuralwatt feat(neuralwatt): add neuralwatt.com charmbracelet/catwalk#247 both did
  this; coralbricks chore: add coralbricks provider charmbracelet/catwalk#502 did not, which looks like an oversight.)
• internal/providers/providers_test.go — TestAIMLAPIDefaultHeaders.

Shape follows the last three provider additions. No cmd/aimlapi/main.go
auto-updater: GET /v1/models returns 353 chat entries, which is not a useful
model picker, and regenerating it would churn the config on every catalog
change. The curated route is the coralbricks (charmbracelet#502) shape.

Why a hand-curated list of nine

353 chat models are reachable. The nine here are the coding-relevant flagships
plus a cheap small model, each one checked live.

Attribution headers

default_headers is the seam the repo already uses for exactly this —
aihubmix.json ships APP-Code: IUFF7106, and openrouter.json,
vercel.json and huggingface.json all ship HTTP-Referer/X-Title.

"default_headers": {
  "HTTP-Referer": "https://charm.land",
  "X-Title": "Crush",
  "X-AIMLAPI-Partner-ID": "part_crush",
  "X-AIMLAPI-Source": "agent/crush"
}

HTTP-Referer and X-Title name Crush, not us — they identify the calling
application, matching the values openrouter.json and vercel.json already
carry. The two X-AIMLAPI-* headers are ours and are scoped to this provider
entry only; nothing else in the database carries them, and
TestAIMLAPIDefaultHeaders asserts that no other provider ever does, so
attribution cannot ride along to somebody else's API or to a proxy fronting
ours.

The partner id is regex-tested against ^part_[A-Za-z0-9]{1,64}$. This matters
because the gateway ignores a malformed value rather than rejecting the
request — a typo would be invisible at runtime and simply earn nothing.

Verification

Build: go build ./... → exit 0.

Lint: gofumpt -l . → clean. golangci-lint run ./... (v2.13.1, the
version pinned in mise.toml) → 0 issues.

Tests (go test ./... -v, subtest counts):

	pass	fail
baseline, pristine checkout of this fork's main	40	1
after this branch	41 (+ TestAIMLAPIDefaultHeaders)	1

The single failure is identical before and after and is not ours:

--- FAIL: TestValidDefaultModels/GitHub_Copilot
    Default large model "claude-sonnet-4.6" not found in provider "GitHub Copilot"

It is already fixed upstream in 155f0c2 fix(copilot): fix default model, three
commits ahead of where this fork's main currently sits.

Model ids: all 9 checked against the live catalog as id or alias, and
filtered to type == "openai/chat-completions" — 9/9 present as canonical ids
on the chat surface, 0 stale. Each was additionally called live: all 9 returned
HTTP 200 for a plain completion, accepted reasoning_effort, and produced a
real tool_calls response.

One real inference call, through the added code path — not curl, not a
mock. A small program imports charm.land/catwalk/pkg/embedded, looks up
InferenceProviderAIMLAPI, and takes the endpoint, the model ids, the API-key
env var name and the headers entirely from the entry this PR adds:

provider name        : aimlapi.com
provider id          : aimlapi
api_endpoint         : https://api.aimlapi.com/v1
api_key env          : AIMLAPI_API_KEY
default large model  : openai/gpt-5.6-terra
default small model  : openai/gpt-5.6-luna
default header       : HTTP-Referer: https://charm.land
default header       : X-Title: Crush
default header       : X-AIMLAPI-Partner-ID: part_crush
default header       : X-AIMLAPI-Source: agent/crush

chat  : HTTP 200  model=gpt-5.6-terra  usage=17/10/27
chat  : content="catwalk aimlapi provider works" finish_reason=stop
tools : HTTP 200  model=gpt-5.6-terra  usage=137/17/154
tools : tool_call=get_weather args={"city":"Paris"} finish_reason=tool_calls
small : HTTP 200  model=gpt-5.6-luna  usage=13/6/19
small : content="small model ok" finish_reason=stop

Server: go run . then GET /v2/providers → 42 providers, aimlapi.com
present with 9 models and all four headers intact through JSON round-trip.

What could not be verified

• reasoning_levels is ["low","medium","high"] for every model. Those
  three were each confirmed to return HTTP 200. xhigh, max, none and
  minimal appear on other providers in this repo but were not tested here, so
  they are not claimed.
• Cache pricing. cost_per_1m_out_cached is the cache read price and
  cost_per_1m_in_cached the cache write price, per CRUSH.md. Write prices
  are published for the Anthropic and GPT-5.6 entries only; the rest are left at
  0 rather than guessed.
• supports_attachments is set from live probes, not from the published
  catalog. For anthropic/claude-opus-5 and anthropic/claude-sonnet-5 the
  catalog's modalities.input says ["text"], but both models in fact accept
  and correctly describe an inline image; deepseek/deepseek-v4-pro correctly
  rejects one with HTTP 400. This is an upstream metadata bug on the aimlapi.com
  side and has been reported there.
• Prices come from GET /v1/models?include=pricing and are a point-in-time
  snapshot. Without a cmd/aimlapi updater they will drift; refreshing them is
  a manual edit.