feat(chat): make per-stage token limits and temperature configurable via agents.yaml

[DarkGenius]

Description

The agentic chat pipeline currently has six hardcoded max_tokens values
(thinking=1200, observing=1200, responding=1800, answer_now=1800,
acting=1500, react_fallback=800) and a hardcoded temperature=0.2 in
AgenticChatPipeline._completion_kwargs. These limits cannot be tuned without
editing source code, and the 1800-token cap on responding truncates long
final responses mid-sentence on real-world prompts (reproduced on a Russian
LaTeX-heavy "VC-dimension proof plan" question — output cut around the 1800-token
boundary, ~4500–4800 chars).

This PR lifts these knobs into data/user/settings/agents.yaml under a new
capabilities.chat block, mirroring the existing per-capability config pattern
(capabilities.solve, capabilities.research, etc.), but with per-stage
sub-sections to give granular control:

capabilities:
  chat:
    temperature: 0.2
    responding:     { max_tokens: 8000 }
    answer_now:     { max_tokens: 8000 }
    thinking:       { max_tokens: 2000 }
    observing:      { max_tokens: 2000 }
    acting:         { max_tokens: 2000 }
    react_fallback: { max_tokens: 1500 }

Defaults are bumped (responding/answer_now: 1800 → 8000) to fix the truncation
bug out of the box. _ChatLimits.from_config provides safe fallbacks if the
yaml block is missing entirely (legacy installs), and coerces malformed values
(strings, non-dict stage entries) back to defaults instead of crashing.

Implementation notes

• New get_chat_params() and DEFAULT_CHAT_PARAMS in
  deeptutor/services/config/loader.py. Kept separate from get_agent_params
  because the chat capability has a nested per-stage shape, while
  get_agent_params returns flat {temperature, max_tokens}.
• DEFAULT_AGENTS_SETTINGS in deeptutor/services/setup/init.py extended so
  fresh installs auto-generate the chat block in their agents.yaml.
• AgenticChatPipeline.__init__ loads the config once into self._chat_limits
  / self._chat_temperature; the six callsites and _completion_kwargs read
  from these instance attrs.

Related Issues

• N/A (internal bug report — chat answer truncation on long responses)

Module(s) Affected

• agents (chat pipeline)
• api
• config (loader + setup defaults)
• core
• knowledge
• logging
• services
• tools
• utils
• web (Frontend)
• docs (Documentation)
• scripts
• tests
• Other: ...

Checklist

• I have read and followed the contribution guidelines.
• My code follows the project's coding standards.
• I have run pre-commit run --all-files and fixed any issues.
• I have added relevant tests for my changes.
• I have updated the documentation (if necessary).
• My changes do not introduce any new security vulnerabilities.

Test plan

• New tests/services/config/test_chat_params_config.py (10 tests):
  • get_chat_params() returns defaults when agents.yaml is missing or the
    capabilities.chat block is absent.
  • Partial overrides (single stage / temperature only) merge correctly with
    defaults.
  • Full round-trip with all stages overridden.
  • _ChatLimits.from_config handles empty dict, DEFAULT_CHAT_PARAMS,
    string-numbers, garbage strings, and non-dict stage values.
• Existing related tests still pass: tests/agents/chat/,
  tests/capabilities/test_answer_now.py, tests/services/config/
  (70 passed locally; nothing depended on the old hardcoded literals).

Additional Notes

• Backwards compatible: existing user agents.yaml files that don't have
  the chat block keep working through DEFAULT_CHAT_PARAMS deep-merge in
  get_chat_params(). Behaviour changes only because the defaults are
  larger, which is the intended fix.
• temperature is now a single value shared by all chat stages
  (matches the existing single-knob pattern of other capabilities).

[pancacake]

Thanks for your contribution!