fix(agent): retry title generation with a larger budget for unconfigured reasoning models

[albatrossflyon-coder]

Fixes #3284.

Root cause

GenerateTitle already bumps the output-token budget for models flagged CatwalkCfg.CanReason (40 → DefaultMaxTokens), specifically so a thinking model doesn't burn its whole budget on hidden reasoning content before producing a title. It also tries a /no_think prompt hack for templates that respect it.

CanReason comes from static model metadata (Catwalk's known-model registry). A hand-configured custom/local provider — like a local llama.cpp server serving a thinking-capable model — has no such metadata unless the user explicitly sets it. If it's unset, the model still reasons by its own chat template default (the /no_think hack only works for templates that specifically respect that convention) — crush just doesn't know to give it room. So it burns the tiny 40-token budget entirely on hidden reasoning, produces no title text, and every attempt (small model, then large model) falls through to the "Untitled Session" fallback — matching this issue's exact symptom.

Fix

Detect this directly instead of depending on correct prior configuration: if a title attempt hits FinishReasonLength and the response contains reasoning content even though CanReason was false, that's a template-agnostic signal the model needs more room regardless of what its config declares. Retry that same attempt once with a larger budget (DefaultMaxTokens, or a 2000-token fallback if that's also unset for a custom model) before moving to the next model attempt.

Extracted the decision into shouldRetryTitleWithLargerBudget so it's directly unit-testable without mocking the full Stream/Agent machinery.

Testing

• Added TestShouldRetryTitleWithLargerBudget (4 cases: unconfigured reasoning model hitting the limit → retry; already-bumped budget → no further retry; hit limit with no reasoning content → not a thinking-budget problem, no retry; finished normally → no retry).
• go build ./... and go vet ./internal/agent/... clean.
• Confirmed no other test in the package broke: go test ./internal/agent/... -run "^$" compiles clean across all subpackages.