perf(ttft): salvage easy-win startup optimizations from #3346

[teknium1]

Summary

Salvages four low-risk, high-value PRs from the TTFT startup optimization investigation (#3346) by kshitijk4poor. Cherry-picked onto current main with authorship preserved.

Cherry-picked commits

PR	Commit	What it does
#3361	5bb3b306	Dedupe tool availability checks — shared check_fn functions called once instead of N times per get_definitions()
#3362	2077bda9	Short-circuit vision auto-resolution — try preferred provider first, avoid probing all backends (~3964ms → ~77ms)
#3363	449a9702	Lazy Claude Code version detection — defer claude --version subprocess to first Anthropic OAuth use, not import time
#3364	36481ec9	Reuse loaded toolsets for skills prompt — derive available toolsets from already-loaded tools instead of re-probing all check_fns

What was NOT included (needs more discussion)

• perf(ttft): move runtime metadata out of cached prompt #3365 — Move runtime metadata out of cached prompt. Changes observable model behavior (Conversation started: label, metadata location). Worth doing but needs careful integration.
• perf(ttft): cache skills prompt metadata across agents and processes #3366 — Cache skills prompt with disk snapshot. Biggest perf win (297ms → 103ms) but introduces code duplication between prompt_builder.py and skills_tool.py. Needs refactor to extract shared logic.
• docs(ttft): record startup benchmark findings #3367 — Documentation only (findings.md). Can merge independently.

Test results

6295 passed, 0 failed (full suite minus MCP).

Co-authored-by: kshitijk4poor 82637225+kshitijk4poor@users.noreply.github.com