feat: add opt-in GCF output format for the table tools

[blackwell-systems]

Summary

Adds an opt-in output format, EQUIBLES_OUTPUT_FORMAT=gcf, that renders the table tools
as a Graph Compact Format generic wire instead of a markdown table.
GCF factors the column names into a single header and drops the per-cell | padding and
the separator row, so the response uses fewer tokens when it crosses the LLM boundary.

It hooks a single seam: MarkdownTable.Render, where the table tools already materialise
their rows. So it covers every tool that renders through it (short data, holdings, insider
trading, congress, FRED, COT, government contracts, FDA, …) with no per-tool changes.

It preserves your formatting exactly

This is the important part: GCF here reuses the exact cells the markdown table would
show. The compact-USD ($1.23B), comma-grouped counts, adaptive-decimal prices, em-dash
nulls, split-adjusted values — all of it is emitted verbatim. GCF only changes the framing
(the | padding and the separator row), never the values a model reads. So the careful
per-column formatting in McpFormat is fully retained; the wire is just denser.

Deliberately conservative

• Opt-in and additive. Default output is markdown, unchanged unless the variable is set.
• Never dropped or garbled. If any row does not split into the same number of cells as
  the header, that result falls back to the markdown table. Data pipes escaped by
  MarkdownTable.EscapeCell are decoded back, not split into phantom columns.
• Lossless. The GCF wire round-trips back to the same cell values.
• Zero new footprint. BlackwellSystems.Gcf is a zero-dependency package (netstandard2.0),
  pinned exact via central package management.

Scope note: this covers the tools that render through MarkdownTable.Render (the single-table
search tools). The few tools that assemble multi-table output with the StringBuilder helpers
stay markdown-only for now; the same seam extends to them later if you want.

Measured on the tools' table shapes

o200k tokenizer, GCF vs the current markdown table, lossless round-trip verified:

• Short-volume style (comma columns): ~15% fewer tokens
• Holdings style (institution + shares + compact-USD value): ~13–14% fewer tokens

The win is the markdown framing (| per cell + the separator row); it scales with row count.

Why GCF, beyond size

The harder question than size is whether a model reads the compact form as accurately as the
table. GCF is designed and evaluated for that:

• Generic-profile comprehension: 100% on every frontier model (Claude Opus/Sonnet/Haiku,
  GPT-5.5, Gemini 2.5 Pro / 3.1 Pro / 3.5 Flash), 27 runs, 11 models, 4 providers; equal to or
  better than JSON on every model. https://gcformat.com/guide/benchmarks
• The grammar was reverse-engineered from attention-level research into how BPE tokenization
  shapes what a transformer can represent, then validated by controlled from-scratch training
  at 1.3B scale, measured causally rather than by an LLM judge. Four papers (2026, under review):
  • GCF: A Token-Optimized Wire Format for Structured LLM Interactions — https://doi.org/10.5281/zenodo.20579817
  • Tokenizer-Attention Coupling — https://doi.org/10.5281/zenodo.20925910
  • Stranded Attention — https://doi.org/10.5281/zenodo.21158886
  • Developmental Atlas of Attention Head Specialization — https://doi.org/10.5281/zenodo.21205389

GCF is already in production across observability, developer tooling, and network automation:
Chrome DevTools MCP (Google) merged it
as an experimental format; Speakeasy (customers include Google, Verizon,
Mistral AI) ships it in their oq CLI; OmniRoute vendored it into its
gateway; NetClaw benchmarked it against TOON on
real network data and replaced TOON; and many others across the MCP and agent ecosystem.

The change

• EQUIBLES_OUTPUT_FORMAT=gcf read in GcfTable, hooked into MarkdownTable.Render.
• BlackwellSystems.Gcf package reference (central version pin).
• Unit tests: the encode + lossless round-trip, the column-mismatch fallback, escaped-pipe
  decoding, formatting preserved verbatim, and that default markdown is unchanged
  (dotnet test green: 4,353 unit tests pass, CSharpier clean).
• README documents the variable.

GCF is open source and MIT-licensed. Format, spec, SDKs, and benchmarks: https://gcformat.com
Happy to adjust the variable name or extend to the multi-table tools; reproducible numbers and
harness on request.

[CLAassistant]

All committers have signed the CLA.