v1.1.0 — async-default verbs + LLM chat (BREAKING)

1.1.0 — ⚠️ BREAKING

This release redefines the inference verbs to make async the default and
adds a normalized LLM chat surface. Adoption was effectively zero, so the
verb flip ships without a deprecation cycle.

Breaking — verb redefinition (SEG-354)

Verb	1.1.0 behavior	Was (≤1.0.x)
run(slug, **params)	async — v2 queue-backed, submit + poll until done; returns the result dict	sync v1, returned httpx.Response
run_sync(slug, **params)	sync — single blocking POST /v1/{slug}; returns httpx.Response	new (this is the old run)
run_async(slug, **params)	removed — no alias; raises AttributeError / ImportError	the old async one-shot
submit_async(slug, **params)	unchanged — returns AsyncJob (.wait/.status/.result)	unchanged
stream(...)	raises NotImplementedError (gateway has no SSE yet)	silently returned None

Migration:

# 1.0.x  ->  1.1.0
segmind.run(slug, ...)          # sync image/etc.   ->  segmind.run_sync(slug, ...)
segmind.run_async(slug, ...)    # async one-shot     ->  segmind.run(slug, ...)
segmind.submit_async(slug, ...) # unchanged

run (async) inherits the 600s job.wait default. For long / video
models the documented path is submit_async() + job.wait(timeout=…), not
run. All model params — including any literally named timeout / interval
— are still forwarded verbatim to the model body (SEG-339).

Added — LLM chat surface + output normalization (SEG-344)

The gateway returns each provider's raw native JSON in three incompatible
shapes (OpenAI choices[0].message.content, Anthropic content[0].text,
Gemini candidates[0].content.parts[0].text). The new chat surface hides that:

• chat(slug, *, messages=[...] | prompt=..., **opts) → ChatResponse — async (mirrors run).
• chat_sync(...) → ChatResponse — sync single POST /v1.
• submit_chat(...) → AsyncChatJob (.wait/.status/.result/.request_id).
• ChatResponse: .text (provider-normalized), .json() (parse JSON-mode output), .usage, .finish_reason, .tool_calls (empty until the gateway forwards tools), .raw (always the native body).
• image_url(url | local_path) — builds an OpenAI image content-part; local paths are base64-encoded into a data: URI (recommended over URL, whose fetch path 400s on some hosts).
• The v2 async output is a stringified copy of the provider JSON; the async chat path json.loads-es it before normalizing.
• stream= kwarg reserved on chat(); raises NotImplementedError until the gateway grows SSE.

100% additive apart from the verb rename above.

1.0.1

What's new in 1.0.1

v2 async inference (SEG-52)

• segmind.submit_async(slug, **params) → AsyncJob handle
• segmind.run_async(slug, **params, timeout=600, interval=1.0) — one-shot submit + poll to completion
• AsyncJob.wait() / .status() / .result()
• New exceptions InferenceFailed (server FAILED) and InferenceTimeout
• Tolerates heimdall's 4xx-on-FAILED responses; derives the v2 endpoint from base_url

SDK traffic attribution (SEG-319)

• The client now sends X-Initiator: SDK-PY, so Segmind can attribute SDK traffic (SDK-PY sync / SDK-PY-V2 async) instead of bucketing it as OTHERS

Maintenance

• CI: bumped GitHub Pages actions past the deprecated upload-artifact@v3

Full test matrix green on Python 3.9–3.12.