feat: add flag-gated TOON compression (~10-39% payload reduction) + validate on Ollama Cloud GLM-5

[Plaidmustache]

Summary

This PR adds TOON-based prompt-context compression behind safe flags and validates behavior on Ollama Cloud (glm-5:cloud).

TOON project: https://github.com/toon-format/toon

What’s included

• TOON compression integration (flag-gated, default OFF)
• Safe fallback behavior (TOON_FAIL_OPEN=true)
• No protocol payload mutation (tool schemas/call envelopes remain unchanged)
• Unit tests covering:
  • flag-off no-op behavior
  • fail-open on encode errors
  • protocol-safety boundaries
• Runtime documentation updates with validation results
• Startup bug fix in budget manager (dbPath scope issue)

Why

• Reduce prompt payload size for large structured JSON context while preserving behavior.
• Measured repeatedly in validation one shot "task planning run" : 6416 -> 5854 bytes (~8.76% reduction), i.e. roughly 9–10%.

Validation notes

Tested on:

• MODEL_PROVIDER=ollama
• OLLAMA_ENDPOINT=http://127.0.0.1:11434
• OLLAMA_MODEL=glm-5:cloud

Observed:

• TOON OFF: no TOON conversion logs
• TOON ON: TOON conversion logs present + byte reduction stats
• Provider routing remained ollama

Why TOON (external benchmark context)

From the TOON project’s published benchmarks:

TOON benchmark snapshot (from TOON docs)

Format	Efficiency (acc% / 1K tok)	Accuracy	Tokens
TOON	26.9	73.9%	2,744
JSON compact	22.9	70.7%	3,081
YAML	18.6	69.0%	3,719
JSON	15.3	69.7%	4,545
XML	13.0	67.1%	5,167

Efficiency score formula: (Accuracy % ÷ Tokens) × 1,000 (higher is better).

Key takeaway from TOON docs: TOON reached 73.9% accuracy vs JSON’s 69.7% while using ~39.6% fewer tokens.

• Reported mixed-structure result: ~39.6% fewer tokens vs pretty JSON, with higher retrieval accuracy (73.9% vs 69.7%).
• Higher efficiency score (accuracy per 1K tokens): TOON 26.9 vs JSON 15.3.
• Designed as a lossless JSON representation (round-trip safe), so app logic can keep JSON while prompt payload is compressed.
• Structure hints ([N] row counts and {fields} headers) are intended to improve schema-following reliability.

Practical caveats (also from TOON docs):

• Deeply nested or non-uniform data may favor compact JSON.
• For latency-critical paths (especially some local/quantized setups), benchmark both formats and choose the faster one.

References:

• https://github.com/toon-format/toon
• https://toonformat.dev/guide/benchmarks

Safety

• Default remains TOON_ENABLED=false
• Fail-open enabled path preserves original JSON on conversion errors
• No personal paths, usernames, or secrets included in committed diff

[veerareddyvishal144]

Thanks @Plaidmustache for this PR.
This works really well.