Summary

• Replace configurable compression thresholds with an input-limit budget derived from context window minus output and safety reserves.
• Add persisted token anchors so compression pressure can reuse provider-reported prompt usage, then estimate only the unmatched tail.
• Remove the old compressionThreshold / compression-start threshold contract from Rust DTOs, frontend API types, event projection, and UI handlers.
• Improve compression diagnostics, final stream usage logging, and compression card saved-token percentage rendering.

Fixes #

Type and Areas

Type:

bug fix / refactor

Areas:

Rust core, agentic session/execution/persistence, desktop/Tauri APIs, event contracts, AI adapters, web UI

Motivation / Impact

The compression trigger now tracks provider-visible request pressure more accurately and reserves response headroom before sending another round. This reduces the chance that local token estimates or stale ratio thresholds delay compression until the request is too close to the model context limit.

The change also removes the old threshold field from internal session/config/event surfaces and keeps the frontend compression UI aligned with the new contract. Compression cards now display saved percentage as the amount reduced, not the remaining compression ratio.

Verification

• pnpm run type-check:web passed.
• cargo test -p bitfun-core token_anchor --lib passed: 7 passed.
• cargo test -p bitfun-core compression_trigger_budget --lib passed: 2 passed.
• cargo test -p bitfun-core auto_compression_pressure --lib passed: 4 passed.
• cargo check --workspace passed.

Reviewer Notes

• Token anchors are stored per session in token-anchors.json and are pruned when context is rewritten, restored, or emergency-truncated.
• Auto compression now triggers when total_tokens >= input_limit, where input_limit = context_window - output_reserve - safety_reserve.
• Default output reserve is 16k tokens when provider max_tokens is unset; safety reserve is 10k tokens.
• This intentionally removes the old compressionThreshold session config and threshold compression-start event payload.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.