Detect Copilot agent mode at chat time and default optimizeFor to "turns

[navintkr]

The optimizeFor knob (tokens | turns | balanced) is exactly
the right trade-off for Copilot agent mode vs one-shot chat. But
right now the user has to set it manually in settings or policy.
We should detect agent-mode-like signals at request time and flip
to turns automatically (with a way to opt out).

Possible signals

• Presence of tools in the ChatRequest.toolReferences array (agent
  mode passes tools; ask-mode doesn't).
• Prompt keywords that strongly suggest multi-step work:
  "refactor across", "migrate", "implement ... and write tests",
  "investigate why", "walk the codebase".
• The task classifier already returns code_large / agentic —
  treat those as an auto-bias to turns regardless of user setting.

What to do

1. In src/participant.ts, inspect
   request.toolReferences on each turn; record whether any tools
   were attached.
2. In src/core/index.ts analyzeAsync, if
   optimizeFor came in as "auto" (new value, default?), resolve
   to "turns" when (a) tools are attached, (b) task is
   code_large or agentic, (c) judge turns estimate > 5.
   Otherwise resolve to "tokens".
3. Surface the chosen mode in the footer (e.g. opt=turns (auto: tools)).
4. Add a setting tokenProctor.optimizeFor default "auto" and
   document the resolution rules.

Acceptance

• New "auto" option in the settings enum + README.
• Agentic prompts (or prompts with tools) resolve to turns.
• Q&A prompts resolve to tokens.
• Footer shows why the mode was chosen.

Why this matters

The #1 failure mode today is a user who installs Token Proctor and
never touches settings — they get the tokens default, which is
actively wrong for their 20-turn agent loops. This fixes the
default-on experience.