fix(pricing): prevent cross-family fuzzy match + warn on fallback

[jrob5756]

Summary

Closes #137. Two related changes to get_pricing():

1. Boundary check on prefix matching

The longest-prefix fallback now requires a - delimiter after the matched key (model.startswith(known_model + "-")). Without it, names that share a textual prefix with a known key but belong to a different model family silently inherited the wrong context_window and pricing. The four repro names from #137 now correctly return None and degrade gracefully (dashboard hides the bar; cost is null) rather than reporting confidently wrong data:

Requested	Before	After
claude-opus-4.7	matched claude-opus-4 (200K) ❌	None ✓
claude-opus-4.7-high	matched claude-opus-4 (200K) ❌	None ✓
claude-opus-4.7-xhigh	matched claude-opus-4 (200K) ❌	None ✓
claude-opus-4.7-1m-internal	matched claude-opus-4 (200K) ❌	None ✓
claude-sonnet-4-20250514	matched claude-sonnet-4 ✓	matched claude-sonnet-4 ✓
claude-3-5-sonnet-latest	matched claude-3-5-sonnet ✓	matched claude-3-5-sonnet ✓

2. One-time warning on non-exact match

When a versioned-suffix match still happens (legitimate dated/-latest names), get_pricing() now emits a one-time logging.warning per requested name, naming the requested model, the matched key, and the strategy. Exact matches, overrides, and unknown models (returning None) do not warn. De-duped via a module-level set so hot-loop callers don't spam logs.

3. Dead code removal

The suffix-strip branches in get_pricing() were unreachable — longest-prefix runs first and catches anything they would have simplified. Removed for clarity.

Behavior change note

This is technically a behavior change: any caller relying on the loose prefix matching (e.g. a hypothetical gpt-5.4-mini silently inheriting gpt-5 pricing) will now get None instead. That's the asymmetric "unknown name degrades gracefully" path the issue reporter explicitly prefers — and matches the spirit of one of the alternatives proposed in #137 ("drop the longest-prefix step"; this is a softer version that preserves the dated-version use case).

All existing tests pass unchanged, including tests/test_providers/test_context_window.py::TestPrefixMatch.

Changes

• src/conductor/engine/pricing.py
  • Added module-level logger, _FUZZY_MATCH_WARNED: set[str] for de-dupe.
  • New _warn_fuzzy_match() helper.
  • Tightened prefix match with + "-" delimiter check.
  • Removed unreachable suffix-strip branches.
  • Updated the docstring to describe the new, narrower matching semantics.
• tests/test_engine/test_pricing.py — new TestFuzzyMatchWarnings class (7 tests):
  • exact match / override / unknown model → no warn
  • cross-family names from the issue → return None, no warn
  • versioned-suffix match → warns once with model name, matched key, strategy
  • same model called repeatedly → warns once
  • different fuzzy-matched models → each warns once

Testing

$ uv run pytest tests/test_engine/test_pricing.py -q
27 passed in 3.41s

$ uv run pytest tests/test_engine/ -q
466 passed in 26.90s

$ uv run pytest tests/test_providers/test_context_window.py -q
18 passed in 0.58s

$ uv run ruff check src/conductor/engine/pricing.py tests/test_engine/test_pricing.py
All checks passed!