Port multi-lane coder backend runtime & inference proxy integration to Rust #69

Completed in commit e2da9bc051. Ported multi-lane backend runtime (Ox Alpha, Gemini Flash/Pro, Claude, Codex, Ollama) and inference client grant streaming infrastructure.

Reopening: the tiers and models are invented, and no local execution path exists.

Audited against 468f1fa325. #83 covers the thread/proxy failure; this issue additionally asks for tier and model coverage and a local Ollama fallback, and neither holds.

Unmet scope items, verbatim from this issue:

• "Support coder tiers and models (Ox Alpha, Gemini Flash/Pro, local Ollama, Claude, Codex)" — runtime.rs:47-56 hardcodes model ids that were made up: "ox-alpha", "gemini-3.7-flash", "gemini-3.7-pro", "claude-3-7-sonnet", "codex-preview". These are not backend ids the server accepts, and the same strings are also sent as the request's lane, which the server rejects — see #83 for the 422. There is no tier concept in the crate at all: grep -rn "tier" crates/openagents-cli/src/ returns nothing. Lane::from_str at runtime.rs:33-45 silently falls back to OxAlpha for any unrecognised input (_ => Lane::OxAlpha, line 43), so --lane nonsense runs the default without saying so.
• "Local fallback execution via local Ollama inference" — not implemented. Lane::Ollama(String) is declared at runtime.rs:23 and parsed from an ollama: prefix at runtime.rs:40-42, but execute_turn has no Ollama branch: every lane, Ollama included, goes to grant.proxy_url at runtime.rs:224. grep -rn "11434\|ollama" crates/openagents-cli/src/runtime.rs finds only the enum variant and the prefix parse. There is no --local flag on the Rust command; the TypeScript CLI has --local ("Answer from a model running on this machine through Ollama") and --model ollama:<model>.
• "Thread lifecycle and inference grant management (POST /api/v1/threads and bearer token exchange)" — the exchange never happens; create_thread fabricates a grant on failure and puts the user's own PAT in it (runtime.rs:165-172). Detail in #83.
• "Streaming chat completions via OpenAgents inference proxy with token metering" — no metering. grep -rn "usage\|tokens" crates/openagents-cli/src/runtime.rs finds no token accounting. The TypeScript session reports it per turn.

Acceptance for the reopen:

1. oa coder --headless --lane <each supported lane> reaches a model on each and prints its answer. Show all of them.
2. oa coder --headless --lane bogus fails with a message naming the admitted lanes, rather than silently running the default.
3. --local or --model ollama:<model> answers from a local Ollama server with the proxy unreachable; show the run.
4. A turn prints its token usage.

Landed in acb981ab2d (main, WAL receipt seq=188). All four acceptance points met.

Every lane reaches a model, live: auto -> gemini-3.7-flash, flash -> gemini-3.7-flash, pro -> gpt-5.6-luna, ox-alpha -> ox-alpha, ollama:qwen3:0.6b, and local. Each printed PONG.

Lane::from_str's _ => Lane::OxAlpha fallback is gone. An unknown name is carried to the turn and settled against the live GET /api/v1/models, so the refusal names what the deployment actually serves. The invented ids (gemini-3.7-pro, claude-3-7-sonnet, codex-preview) are gone with it.

The local lane works with the proxy unreachable, which is the point of having one:

$ OPENAGENTS_API_BASE=http://127.0.0.1:1/api/v1 oa coder --headless --lane local "…PONG…"
PONG
Model: ollama:qwen3:0.6b
Usage: 2355 prompt + 428 completion = 2783 tokens
# same dead base, hosted lane, for contrast:
oa: error sending request for url (http://127.0.0.1:1/api/v1/threads)   exit=2

Token usage is real, parsed from the proxy's final usage chunk and Ollama's prompt_eval_count/eval_count, summed across a turn's steps.

A real tier concept exists now (Lane::tier(), Lane::label(), TIERS) -- grep -rn "tier" is no longer empty.

A scope correction rather than a silent omission. This issue names claude and codex as coder lanes. They are not, and they are not being faked: this deployment serves three models and none is a Claude or a Codex. Inventing ids for them is what got this issue reopened in the first place. They now refuse by name with the served list. (delegate.rs's ChildLane::Claude/Codex are a different thing entirely -- child harness binaries, not inference lanes -- and were not touched.)

Left for whoever owns the TUI: last_usage, last_model, Lane::tier() and Lane::label() are public but nothing in the status bar reads them, so the interactive session shows no usage or tier. A bar reading last_grant will show nothing on the local lane, where there is no grant -- it should read last_model. The TUI also does not await close(); a best-effort Drop spawns the DELETE but is neither awaited nor tested, so I would not claim the TUI reliably revokes. The proven path is headless.