…model

- Add ModelInfo struct and model state fields to AcpConnection
- Parse available models from session/new response
- Add session_set_model() and resolve_model_alias() methods
- Add !model Discord command to list and switch models
- Support aliases: opus, sonnet, haiku, auto

Kiro ACP session/set_model verified on Kiro CLI v1.29.3

Move the entrypoint.sh logic into main.rs so openab is independent of any
external shell wrapper. Before loading config, it now:

  1. Restores ~/.local/share/kiro-cli/data.sqlite3 from KIRO_CRED_B64 if
     missing (idempotent — skips when file already exists, e.g. from a
     mounted volume).
  2. Generates the config file from DISCORD_BOT_TOKEN / DISCORD_CHANNEL_ID
     env vars when the path doesn't exist. Secrets are written as literal
     ${VAR} placeholders so they get expanded by load_config() at read
     time and never land on disk in plaintext.

Why: Zeabur deployments were intermittently running images built without
the entrypoint shim (e.g. when the source branch was switched mid-flight),
producing "failed to read /etc/openab/config.toml" with no entrypoint logs.
With the bootstrap inside the binary, the deploy is robust to any image
build path that happens to omit our shell wrapper.

The Dockerfile entrypoint.sh remains as a belt-and-suspenders no-op when
present.

…plete

Discord's slash commands give a much better UX than text commands:
- Typing /m surfaces /model in Discord's native command picker
- Autocomplete on the model option shows the live list of available
  models (and aliases) with arrow-key selection
- Discord renders the response as an interaction reply instead of
  cluttering the channel as a regular bot message

Implementation notes:

- SessionPool gains a cached_models snapshot updated on every
  session_new(). Autocomplete must respond within Discord's 3-second
  deadline, which rules out spawning a fresh kiro-cli (~10s cold start).
- main.rs adds a background warmup task that creates one session at
  startup so the cache is populated before the first user invocation.
- /model is registered as a guild command on ready() for instant
  propagation (global commands take up to 1 hour).
- The interaction handler defers the response on the set path because
  cold session creation can exceed Discord's 3s window for non-deferred
  replies.
- README documents the applications.commands scope requirement; without
  it the slash command will not appear in Discord clients.

The !model text command is removed entirely — /model is its replacement,
not an addition.

…omplete

- Add interaction_allowlist_ok() helper (channel, thread-parent, user checks)
- Apply to handle_model_command: reject with ephemeral reply if not allowed
- Apply to handle_model_autocomplete: return empty choices if not allowed

Addresses review feedback from thepagent on PR openabdev#186

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Extract pure allowlist logic into ctx-free allowlist_decision() so it can be
unit-tested without a Discord runtime. interaction_allowlist_ok() becomes a
thin wrapper that resolves the thread parent via HTTP and delegates.

Covers: empty lists, channel in/out, thread parent in/out, user in/out, and
the "thread-allowed + user-denied" combination.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>