refactor(provider)!: rig-free provider stack, native Bedrock, minimal deps; refreshed scale claims

[shadaj]

Remove rig from the core provider/agent stack. infinity-provider-protocol now owns a
minimal model API; the Bedrock provider talks to the official AWS SDK directly
(eliminating the maintained rig-bedrock patch); rig survives only as an optional
bridge crate. Re-measured the memory benchmark and refreshed the landing/README claims.

infinity-provider-protocol (rig-free, minimal core)

• Featureless dependency tree is just serde, serde_json, thiserror, async-trait,
  futures-util — no HTTP stack, no LLM SDKs, no tokio.
• Own message types with JSON byte-compatible with the rig 0.31 types they replace,
  locked in by literal-JSON serde tests (persisted histories keep loading). Cleanups:
  OneOrMany → Vec, dropped always-unused fields (ToolCall.signature/
  additional_params, assistant id, Image.detail, audio/video/document content).
• Own completion types: CompletionRequest trimmed to the five fields we use, and a
  plain ModelStream = Stream<Result<StreamChunk, CompletionError>> replacing rig's
  streaming machinery; ModelProvider::invoke_model returns it directly.
• Feature flags keep extras out of default builds:
  • mock — channel-driven mock completion model (replaces the rig-mock crate). A
    self dev-dependency enables it for the crate's own tests, which also removed the
    duplicate inline mock that remote.rs tests previously needed. Downstream tests
    dev-depend on the crate with features = ["mock"].
  • remote — the out-of-process transport and all its deps (tokio, tokio-util, uuid,
    tracing, async-stream, futures-util/sink), gated #[cfg(all(unix, feature = "remote"))] so a Windows transport can join behind the same flag later. Wire
    simplification: requests are the serializable CompletionRequest; chunks travel
    as Result<StreamChunk, String>.

infinity-provider-bedrock (native AWS SDK)

• Rewritten on aws-sdk-bedrockruntime converse_stream; rig-bedrock-patched and
  the root [patch.crates-io] are deleted. Same model catalog (incl. the
  claude-opus-4-6-v1:1m picker entry mapping back to the real model id), params
  merge (caller wins), per-model max_tokens defaults, and the prompt cache point
  appended to the last message (the reason the patch existed).
• Request assembly factored into a pure prepare_request so policy behavior is
  unit-testable without an AWS client; invoke_model is wiring + .send().
• Improvements: mid-stream errors surface instead of silently ending the stream;
  Usage.cached_input_tokens reports cache_read_input_tokens; SDK errors carry the
  underlying service message (agent-core's retry string-matching still works).

New optional crate

• infinity-provider-rig: bridge exposing any rig CompletionModel as a
  ModelProvider, for out-of-tree rig backends. The in-tree stack does not use it.

Consumers migrated

agent-core, daemon, cli, lambda, mcp-bridge, rap-bridge: new import paths, vec![…]
for OneOrMany::one, Option-returning .first(), de-genericized
CompletionEvent/CompletionAction (Done(FinalResponse)), lambda's
aws_config::load_from_env deprecation fixed. Insta snapshots updated (no more
"id": null / "signature": null noise).

Tests kept intentional

Pruned tautological tests (implementation-mirror conversions, serde's own
unknown-field behavior, trivial defaults); kept contract tests: persisted-JSON
formats, Bedrock validation rules, cache-point placement, params/max_tokens/model-id
policies, reasoning round-trip, and the socket-transport integration tests.

Benchmark + landing refresh

• Memory improved: re-ran agent_scale with the published configuration
  (AGENTS=80000 TURNS=20 WAVE=4000, same AMD EPYC 9R14 class host): 8.41 GB @ 80k
  agents vs 8.89 GB before (~103 KB/agent vs ~108; 76k agents measured at 8.0 GB).
  Baseline main reproduced the published curve, so the delta is from this refactor
  (dead-field removal outweighs the extra Vec in tool results — no tinyvec needed).
• docs/src/components/MemoryChart.tsx DATA refreshed (chart now annotates
  "80,000 agents in 8.4 GB", "≈ 103 KB per agent").
• README + landing lead rewritten: "An open-source Rust framework for building
  massively concurrent agentic systems, light enough to fit 75k agents in the memory
  of a Raspberry Pi." (was "fifty/seventy thousand"); scale chapter now titled
  "75k agents on a Raspberry Pi" with ~103 KB per agent.

Validation

fmt, clippy -D warnings (incl. --features e2e-web), full workspace test suite,
Playwright web e2e (7/7), docs prettier clean, THIRD-PARTY regenerated (rig-bedrock
subtree removed, aws-sdk-bedrockruntime added).

BREAKING CHANGE: infinity-provider-protocol types replace rig types throughout;
ModelProvider::invoke_model returns ModelStream; the remote module and mock are
behind the remote/mock features; the provider wire Chunk frame shape changed
(daemon and provider binaries deploy together, as they already do); the rig-mock
and rig-bedrock-patched crates are gone.

[cloudflare-workers-and-pages]

Deploying infinity with    Cloudflare Pages

Latest commit:	f58c02c
Status:	✅  Deploy successful!
Preview URL:	https://6f983505.infinity-dc7.pages.dev
Branch Preview URL:	https://sandbox-25e5957d-5b39-4699-9.infinity-dc7.pages.dev

View logs

[akainth015]

Partial review, I haven't yet finished because it's a massive diff 😅

[akainth015]

It's not clear to me why these changed, though the fields are null so we may not care.

[akainth015]

ToolResultContent::Text(t) is now a type that needs to be .clone()'d? &str?

[akainth015]

ReasoningDelta renames the reasoning field?