[OMEGA-350] token budget fix, V2 implementation

[paul-v-snet]

Description

A ⚠️vibe-coded⚠️ implementation of #337 for Omega V2.

What I did by myself in this PR:

• Analyzed [OMEGA-374] Omega V2: using tools API fields to pass the list of tools to LLM #349
• Prepared plan for rewriting [OMEGA-350] reasoning takes the whole token budget #337
• Iteratively asked the agent to fix/improve things that didn't look good to me
• Did a review
• Built from the local branch, ran live tests, reproduced all the problems that should be fixed, and checked whether they were actually fixed

What's different compared to #337:

• Implemented LLM_TRUNCATED_CALL_HINT, which is added to the LLM error response when a tool call is broken and the finish_reason is length or max_output_token, so that the LLM can try to use a smaller response and solve the issue itself.

How Has This Been Tested?

Ran live tests for OpenAI, ASICloud, and OpenRouter - all passed.

Checklist

• PR contains autogenerated code
• Self-review completed
• Test scenarios above are passed with the version of the code from PR

[TossSky]

@paul-v-snet @vsbogd tested e3a8a38: built the image, ran this branch's CI, and ran live sessions on OpenRouter with z-ai/glm-5.2 at the default config and at maxOutputToken: 120.

What I checked:

• [LLM_USAGE] is at INFO on every call: 105 calls in the 120 run and 58 at the default.
• 17 replies at 120 and 2 at the default came back with finish_reason=length, and each one produced the empty-reply notice.
• The V1 history problem does not exist here. The notice never reaches the prompt: 0 copies across all 105 and 58 REQUEST: lines, because history goes through the episodic buffer where the assistant message is kept as [TOOL CALL] and the send result as RETURN: None. On #337 the same scenario put 45 copies into a single prompt.
• LLM_TRUNCATED_CALL_HINT did not fire in either run, so that path is still unverified live.

CI on this branch: MeTTa tests green, tests/pytest.sh 65 passed, Phase 1 2 failed, 127 passed. The two failures are mock/test_openclaw_delegate_mock.py::test_delegate_stays_async_under_a_slow_gateway_mock and mock/test_pin_invisible_within_iteration_mock.py::test_pin_invisible_within_iteration_mock, and the base branch run fails with exactly the same two, so they are not from this PR.

One thing outside this diff. At the default budget the container died while the agent was writing the escaped JSON of step 6:

ERROR: [Thread main] /PeTTa/src/main.pl:23: user:main Syntax error: Parse error in form: (append-file "/tmp/data.txt" "{\"path\":\"C:\\Program Files (x86)\\AMD\", ...

The tool call is parsed before it reaches eval in src/loop.metta, so a malformed argument takes down the interpreter rather than one iteration. I saw it once and my attempt to reproduce it failed for an unrelated reason, so treat it as a single observation. It deserves its own ticket.

Verdict: PASS for the changes in this PR.