fix(research): derive context budget from actual model context window

[Meur3ault]

Problem

research_tool.py hard-coded the research sub-agent's context budget as
two module-level constants:

_RESEARCH_CONTEXT_WARN = 170_000  # 85% of 200k
_RESEARCH_CONTEXT_MAX  = 190_000            

These values assumed every research model has a 200k context window.
In practice, claude-sonnet-4-6 (the default research model for
Anthropic/Bedrock sessions) has a 1,000,000-token window, so the
sub-agent was being force-stopped at 19% of its actual capacity.

Two secondary issues were also present:

• The warn threshold was 85% of 200k (170k), but the system message
  injected at that point reads "You have used 75% of your context
  budget" — the percentage in the message didn't match the threshold
  that triggered it.
• For any model with a context window smaller than 190k, the hard-coded
  _RESEARCH_CONTEXT_MAX would exceed the API limit, causing a
  ContextWindowExceededError mid-research.

Why context depth matters for the research agent

The research sub-agent's value comes from thorough, multi-step literature
crawls: finding anchor papers, traversing citation graphs, reading
methodology sections, cross-referencing datasets, and pulling working
code from GitHub. A single deep research turn can accumulate dozens of
tool outputs — full paper sections, dataset schemas, code files — before
producing its final summary.

Cutting the sub-agent off at 190k tokens means it is routinely forced to
summarise before it has finished gathering evidence, producing shallower
findings than the model is actually capable of. With the correct 950k
budget, the sub-agent can complete its full citation crawl before being
asked to wrap up.

Fix

Move the constants inside research_handler and compute them from the
research model's actual context window, using the same
_get_max_tokens_safe helper the main session already uses:

_ctx_max = _get_max_tokens_safe(research_model)                               
_research_context_warn = int(_ctx_max * 0.75)  # matches "75%" in the injected
 prompt                                                                       
_research_context_max  = int(_ctx_max * 0.95)  # hard-stop at 95%

The warn ratio is corrected to 75% to match the existing system prompt
text. The hard-stop moves to 95% of the model's real ceiling.

Impact

Model	Old MAX	New MAX
claude-sonnet-4-6 (default)	190,000	950,000
claude-opus-4-7	190,000	950,000
128k-window model	190,000 (exceeds limit)	121,600 (safe)

Tests

Added tests/unit/test_research_context_budget.py (4 cases):

• _get_research_model routing for Anthropic / Bedrock / other models
• research_handler calls _get_max_tokens_safe with the research
  model id, not the main model id

[Meur3ault]

Format checks have passed:
uv run ruff check . uv run ruff format --check .