fix(responses): read-only proxy tools no longer drive the compress loop (炸锅)

[ranxianglei]

Problem

Reported in dog/billion-context-pi#22 (log: `bili (5).log`).

When the model called acp_status (or search_context), the proxy treated it like compress/decompress: it executed it server-side and re-requested the upstream. A model that called acp_status round after round hit the round cap:

```
[acp-diag] round 5 allCalls=[acp_status] ...
[acp-proxy] compress loop limit (5) reached, forwarding completion as-is
```

and the whole turn was discarded — the client received an empty completion (炸锅).

Root cause

PROXY_TOOL_NAMES lumps all 4 proxy tools into one set, but they are not equivalent:

tool	kind	should loop?
compress	mutating	✅ re-request
decompress	mutating	✅ re-request
search_context	read-only	❌ surface + end turn
acp_status	read-only	❌ surface + end turn

A read-only round has no state change to feed back to the model, so re-requesting just makes the model repeat itself until the cap fires.

Fix

Split PROXY_TOOL_NAMES into:

• `MUTATING_PROXY_TOOLS = { compress, decompress }` — drive the re-request loop

• `READONLY_PROXY_TOOLS = { search_context, acp_status }` — execute once, surface the result to the client, end the turn

• Only a round containing a mutating tool triggers an upstream re-request.

• A read-only-only round executes the tool(s), emits the visibility marker, and completes.

• A mixed round (e.g. compress + acp_status) still re-requests, because a mutating tool is present.

This mirrors the Pi adapter (billion-context-pi), where the host executes acp_status once and drives the next turn itself — so it can never loop or discard.

Scope

This PR fixes the Responses path (src/compress-loop-responses.ts), which is what the failing log uses. The Anthropic (src/compress-loop-anthropic.ts) and OpenAI-chat (src/compress-loop.ts) loop files have the same structural bug and should get the same split in a follow-up.

Tests

Two new regression tests in `tests/proxy-responses-stream.test.ts`:

1. read-only-only round — model calls only acp_status → asserts `fetchCalls === 0` (no re-request), `[ACP]` marker present, `response.completed` present, no `compress loop limit`.
2. mixed round — `compress` + `acp_status` → asserts exactly one re-request (mutating tool still drives the loop).

```
ℹ tests 214 (+2)
ℹ pass 214
ℹ fail 0
```

npm run typecheck ✅

Not fixed here (separate issues from the log)

1. Anthropic / OpenAI-chat loop files — same readonly-loop bug, follow-up.
2. Tag echo (~57×) — model regurgitates input <acp tokens=…> ref tags in output; `server.ts` detects+warns only, no strip on the responses path. Cosmetic.
3. reasoning preamble accumulating to 50 — likely v0.1.27 residue (master routes reasoning as a tracked BiliMessage via PR#75); verify on v0.1.31.

🤖 PR opened by agent. Merge is human-only.