docs: define colibri as an experimental inference engine

[ZacharyZcR]

Summary

• define colibri as an experimental inference engine and research platform
• state the objective: optimize the complete inference path to reduce hardware dependency and inference cost
• distinguish measured findings, conditional wins, correctness properties, and open hypotheses instead of presenting every technique as guaranteed acceleration
• document the core techniques: hierarchical placement, learned weight staging, I/O overlap, heterogeneous execution, compressed state, and measured speculation
• publish an open research matrix covering routing history, multi-SSD I/O, automatic hardware planning, representation/compression, speculation, and CPU/GPU overlap
• give contributors a reproducible experiment protocol and explicitly welcome controlled negative results
• update the roadmap and correct the absolute GLM-5.2 int4 think mode: stochastic reasoning loops / EOS starvation on stock v1.0.0; #307's TEMP=0.9 NUCLEUS=0.95 stopgap makes it severe #455 wording
• align the English, Simplified Chinese, Traditional Chinese, and Italian READMEs

Scope

README positioning and research documentation only. No runtime, build, API, or configuration behavior changes.

Validation

• git diff --check
• verified the positioning, measured-findings, and open-experiments sections in all four language files
• reviewed modified Markdown links, tables, and structure

[ZacharyZcR]

Added the Discord community link to the top navigation and contribution section in all four README languages: https://discord.gg/fpQxKnRb

[monotophic]

Authored by Fable 5 in Claude Code, analysis in partnership with @monotophic (We/our = the joint analysis; me/my = @monotophic's own hardware and commitments).

Glad to see this merged — the measured / conditional / correctness / hypothesis taxonomy is the repo formalizing what its best threads already do, and we want to put two concrete datapoints behind it plus a flag of interest on one research-matrix row.

The taxonomy does real work — two cases from our own record

A conditional win, correctly filed: COLI_METAL_RESSET=1 on my M5 Max took decode from 0.89 to a flat 2.28–2.39 tok/s (measured both ways, #379) — a >2.5× win that is platform-conditional by mechanism, not a guaranteed acceleration. Under the old framing that number invites over-claiming; under this taxonomy it files cleanly as a conditional win with a stated condition. That's the framework earning its keep.

A controlled negative result, welcomed: we built and instrumented three generations of prefetch on the storage tier and refuted all three with the instrument (full record on #387; occupancy-vs-saturation trap documented on #537). Under "every technique is an acceleration," that work looks like failure; under this document it's exactly the contribution the protocol asks for. A repo that welcomes scored refutations gets honest numbers.

The "automatic hardware planning" row: we're active in this lane

The research matrix names automatic hardware planning as an open direction. We've been building the measurement-first end of it: a digital twin of Colibri decode — a serial-sum model (disk + expert-matmul + attention + lm_head + other) whose every parameter carries measured provenance and an uncertainty class, validated by hindcast against a 34-run corpus on my M5 Max. Components are atomic and substitutable, so the same model prices a candidate config — or candidate hardware — before anything runs.

What makes planning tractable is that some parameters turn out to be class-constants rather than machine-specifics. Two measured examples from my fleet: sustained STREAM lands at 44% of the spec-sheet bandwidth on two different SoC families (114 of 256 GB/s theoretical on Strix Halo, 120 of 273 on GB10 — best-of-10 triad, working set ≫ LLC), and the ~74% expert cache-hit ceiling on my M5 Max behaves as a routing-concentration constant of model+workload rather than a cache-size effect (#537). Constants like these are what let a planner generalize beyond the box it was calibrated on.

This lane also composes with #673: coli tune searches the config space empirically on one machine; a twin predicts across machines; profiles and predictions cross-validate each other. We've offered a twin-pair profile-transfer experiment on that thread that tests exactly this seam.

Flagging interest, not claiming the row: our methodology and fleet dataset (three machine classes measured across five boxes, more standing up) are available to whoever works this direction, and we expect to bring a concrete proposal here once our current transport-layer work (#671's follow-on rungs) clears the queue.