Unified transcript resume state + same-provider reasoning fidelity

[ryanbbrown]

Summary

This branch replaces the previous provider-specific resume payloads with a single provider-agnostic neutral transcript, then layers same-provider reasoning fidelity on top of it and refreshes the docs/site to match.

• Unified transcript resume state (286390c): built-in sessions keep native in-run state plus a parallel neutral transcript; dump_state returns the transcript so resume state is kind="transcript", provider-agnostic, and strictly validated.
• Tracing from transcript deltas (707b86c): trace projection is derived from transcript deltas (new projections.py) rather than provider-specific plumbing, consolidating tracing.py.
• Same-provider reasoning fidelity (54262c5): native reasoning is preserved across same-provider resume — Anthropic signed thinking, OpenAI encrypted_content, OpenRouter reasoning_details — and degrades to a leading <thinking>-tagged text block on cross-provider resume. Resume state bumped to version=3.
• Docs + behavior contracts (ec73af7): behavior.md RESUME-3/RESUME-7 and docs.md document the resume + reasoning behavior.
• Docs/site refresh (this commit): re-measured LOC comparison table (ThinHarness now 8,658), new LongMemEval-V2 example page alongside the Web Research transcript, explainer gains a reasoning-fidelity subsection in the resume deep-dive, and the README feature-isolation claim was corrected.

Notes

• resume_state can now carry encrypted reasoning blobs and signed thinking — treat it as sensitive, consistent with local traces.
• The about page is generated from the README table via scripts/build_site.py (--check passes).

🤖 Generated with Claude Code

https://claude.ai/code/session_01PFBTUTAZRRH4RtRkLnJWDv