eliziff makes Codex a serious subscription-backed assistant

`codex exec` respawned the CLI every turn (~320-930ms of process init) and
emitted the assistant message only as a single item.completed lump, so
firstContentLatencyMs landed within ~0.4s of totalLatencyMs - users waited out
the whole generation before seeing a character.

Adds a `codex app-server` adapter: one persistent JSON-RPC child per auth
identity, newline-delimited framing, real `item/agentMessage/delta` streaming,
per-thread mike_runtime MCP bridge config, turn/interrupt-backed aborts, and a
transparent fallback to the exec path (BEAVER_CODEX_EXEC=1 forces it).

The app-server has no --ignore-user-config, so it runs against an isolated
CODEX_HOME that mirrors the CLI login; that keeps the operator config.toml MCP
servers, hooks, and plugins out of Beaver turns as the exec path already does.

Measured on codex-cli 0.145.0, gpt-5.3-codex-spark, effort low:
exec 1 content delta at 3.4-4.9s; app-server 8-9 deltas at 2.0-2.9s.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QUsMXzBsEJ27DjJVPSfW5o

Reasoning summaries stream to the client and persist as transcript
events, so falling back to exec after one has escaped would duplicate
the thinking block in the saved chat. Also verified live under review:
two concurrent turns on the shared app-server singleton show no
cross-talk and no truncation (every notification carries a required
top-level threadId per the generated 0.145 protocol schema).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QUsMXzBsEJ27DjJVPSfW5o

MIKE_CODEX_BASE_INSTRUCTIONS replaces the Codex base prompt and
MIKE_CODEX_FEATURES_OFF disables tool-suite features via thread/start
overrides. Measured 15.8k -> 9.7k input tokens/turn on the ChatGPT
backend; the rest is server-side and not removable.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yXxuitxX5h4wm1v88hnga

Direct Responses calls to chatgpt.com/backend-api/codex with OAuth borrowed from the Codex CLI (flow ported from simonw/llm-openai-via-codex, MIT). 24-45 input tokens/turn vs ~16k through the harness; measured live. codex: app-server route unchanged for chat.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Claude-Session: https://claude.ai/code/session_016yXxuitxX5h4wm1v88hnga

Delete the app-server bridge, exec adapter, and MCP tool bridge; codex:
models now stream through codexApi (native function calls, ~50 tokens
overhead vs ~16k). Drop codex-exec legacy id, codex-api: alias, and the
chat isCodex special-cases. Verified live: legalbench abercrombie 84.0
(parity with harness sweep) and a native tool-call round trip.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yXxuitxX5h4wm1v88hnga

The experiment measures LAB harness vs Beaver harness with the model
held constant per pairing (sonnet-4-6 via headless Claude Code; desktop
qwen via ollama). Codex rows are pilot history - no codex in any
experiment cell. Claude-Code-as-product runs relabeled cc-harness-* as
a side observation. AGENTS.md gains the .CMD-shim argv trap and the
corrected design note.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018bhP89cYSdEEjHNPnxacns

server_is_overloaded killed five benchmark turns today; the Responses
transport now retries an attempt up to twice with linear backoff, but
only while nothing from that attempt has reached the caller - a replay
after emitted deltas or tool-call starts would duplicate output. The
persistent-thread response id rolls back to its pre-attempt checkpoint
on retry.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AgUm5EcRKbT3duVFMbrQD

Measured 1,502/2,008 rate-limit errors at c=12 (short low-effort calls
push request rate far above the rerank arms' profile); retry up to 5x
with jittered backoff, error rows re-run via resume.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AgUm5EcRKbT3duVFMbrQD