fix(acp): route codex -c overrides through CODEX_CONFIG for npm adapter

[stephanbrez]

Fixes #27, and adds custom model override support (requested in the #27
comments, not part of the original issue).

Problem (#27)

The npm @agentclientprotocol/codex-acp adapter (v1.1.0+, verified against
1.1.2) does not parse -c overrides from argv — it reads its startup
configuration from the CODEX_CONFIG env var (JSON), parsed at
src/CodexAcpClient.ts and spread unchanged into codex's thread/start
config. As a result, the -c flags appended by _augment_acp_command
(sandbox_mode, approval_policy, model_reasoning_effort) were silently
discarded — sandbox, approval policy, and reasoning effort never reached the
worker.

Fix

Build CODEX_CONFIG in _acp_subprocess_env in three layers, each overriding
the previous:

1. User-supplied CODEX_CONFIG (ambient env).
2. -c overrides parsed from the augmented command string, via
   _parse_codex_c_overrides (handles both double- and single-quoted values).
3. Zenith's three safety keys (sandbox_mode, approval_policy,
   model_reasoning_effort) — always win, since sandbox/approval are
   autonomy-safety requirements and effort is the per-role resolved value.

The -c flags remain in argv for -c-honoring codex-acp builds (harmless if
ignored by the npm adapter). _acp_subprocess_env now accepts
reasoning_effort and acp_command; both call sites (run_node,
run_terminal_review) thread role_config.worker_reasoning_effort and the
augmented command through.

Additional: custom model override

Routing all -c overrides through CODEX_CONFIG also lets a user-specified
codex model reach the adapter — previously dropped for the same reason. A
model supplied via ZENITH_*_ACP_COMMAND (-c model="...") or ambient
CODEX_CONFIG is now preserved (layers 1–2) while zenith's safety keys still
override (layer 3). This is the model-override capability requested in the #27
comments; it is not part of the original issue but rides on the same mechanism.

Verification

Verified end-to-end against the installed adapter: with a CODEX_CONFIG
carrying model, sandbox_mode="danger-full-access",
approval_policy="never", and model_reasoning_effort="max", the adapter's
startup log (app-server.log) shows codexConfig with all four keys received.

Tests

• _parse_codex_c_overrides extracts model + sandbox from double- and
  single-quoted -c flags; empty when no flags present.
• CODEX_CONFIG JSON present with correct effort; defaults to xhigh.
• User CODEX_CONFIG merged (preserves ambient model); command-string model
  wins over ambient model.
• Custom model from augmented command preserved alongside zenith's safety keys.
• Zenith's safety keys override user-supplied safety values; invalid user JSON
  replaced. Claude path has no CODEX_CONFIG.