v0.34.0

Changelog

• c83f23b chore(memory): capture 2026-07-04 — authoritative-meter prediction, tiktoken, 17 providers, optimizer honesty
• 0a6fb01 chore(memory): capture session 7 — 13 providers, opencode reader, v0.33.0, mnemos
• 6bbd8d1 docs(agents): 2 new Known Failure Modes; mark read-guard cross-agent fixed-in-code
• 079c515 docs(agents): 3 new Known Failure Modes (goreleaser transfer 307, read-guard cross-agent, schema forward-only)
• f30e9bc docs(changelog): v0.34.0 — vendor-meter predictions, exact tokenizer, 17 providers
• 01a401f docs(pricing): override example + drift note for the multiplexers
• d8abf7b feat(plans): authoritative window in plan_headroom; harden line-dedup
• c07cdd0 feat(plans): monthly authoritative quota for plan_headroom (Copilot/Cursor)
• d4cacb7 feat(plans): predict from the vendor's own rate-limit meter, not a message count
• fa768c0 feat(providers): add local runtimes + gateways (Ollama, LM Studio, LiteLLM, Vercel)
• d97d6c5 feat(tokenizer): exact tiktoken BPE tokenizer for OpenAI
• e87fd2c fix(optimizers): honest savings estimate + accurate self-descriptions
• 8d8add0 fix(read-guard): scope ledger per agent-context, not per session