Run Laya decisions fully in-process on ONNX Runtime - #55

Adds a new package that runs a local decision model returning typed
choice/score/yes-no answers with full probability distributions in a
single forward pass, rather than generating text.

Backed by Laya (Apache-2.0) over the Jev System One wire protocol,
reached through ElBruno.AI.Jev 0.6.0. The wire contract was validated
end to end against a live laya-serve instance: all three answer types
round trip, the request omits `model` so Laya auto-routes, and extra
Laya fields survive in RawRepresentation.

API:
- IDecisionClient with ChooseAsync / ScoreAsync / AskAsync, plus
  EvaluateAsync for batching several questions into one forward pass.
- DecisionRequest builder (immutable) and DecisionResult reader.
- ChoiceResult.ChoiceOrNull(threshold) so ambiguous inputs can fall
  through to a human instead of being confidently misrouted.
- ProbabilityResult exposes the probability, not just a boolean, and
  AtThreshold re-thresholds without re-running inference.
- AddLocalDecisions() DI registration, and an IJevDecisionClient
  constructor overload for testing without a server.

DecisionOptions.Endpoint is restricted to loopback so prompts cannot
leave the machine through a configuration mistake.

Laya's public checkpoints report uncalibrated confidence, and Laya
itself warns about out-of-range temperatures at startup. The guide,
README and sample all state plainly that DecisionThreshold must be
fitted on the user's own labelled data.

Also registers the package in Directory.Build.props sibling versioning
and adds its pack line to publish.yml.

46 new tests. Full solution: 1998 passing.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

The Decisions package required a running `laya-serve` Python process reached
over loopback HTTP through the ElBruno.AI.Jev SDK. That put a Python install,
a second process and a port in the path of every decision, which is a poor fit
for a library whose whole premise is local in-process inference.

Replace the transport with pure managed .NET: Microsoft.ML.OnnxRuntime for the
forward pass and Microsoft.ML.Tokenizers for the input. `LayaDecisionClient`
becomes `LayaOnnxDecisionClient`. The package was never published, so no
consumer is broken.

Reproducing Laya's numbers required porting four pieces exactly, since the
model scores the hidden state at each `[MASK]` marker and a single extra or
missing token shifts every later marker and silently changes the answer:

- `LayaTokenizer` rebuilds the checkpoint's ByteLevel BPE from its own
  `tokenizer.json`. `BpeOptions.ByteLevel` only does byte<->unicode mapping and
  does not pre-tokenize, so `ByteLevelPreTokenizer` supplies the GPT-2 split;
  without it leading spaces vanish. HuggingFace also matches added tokens
  before BPE runs, including `[MASK]`'s `lstrip`, so that pre-pass is ported
  too. Validated at 22/22 against transformers.js.
- `LayaSequenceBuilder` ports `build_sequence` and `render_options`, keeping
  Laya's budget order: options, then instructions, then the state.
- `LayaRuntimeConfig` reads the fitted temperatures and clamps them into
  Laya's own 0.5-5.0 range. The English checkpoint ships `choice:11+ = 0.1006`,
  which sharpens the logits roughly tenfold and would report a 24% answer as
  99% confident.
- `DecisionMath` divides by temperature then softmaxes, shifting by the maximum
  so a large logit over a small temperature cannot overflow to NaN.

Verified end-to-end: `billing 0.9696 / technical 0.0152 / sales 0.0152`,
matching the reference checkpoint's published model card exactly, and the
sample routes all three tickets correctly with four batched questions each.

DecisionOptions drops Endpoint/ApiKey/Timeout/Model for ModelRepository,
ModelPath, CacheDirectory and IntraOpNumThreads, so the community ONNX export
used is fully configurable.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

The default model source was inferenceprince/laya-onnx, a third-party
repository that could be renamed or deleted out from under consumers.
elbruno/laya-onnx now mirrors that export, verified byte-identical: all
six files match by git/LFS OID, and inference over identical inputs
produces a zero delta against the original.

The weights are unmodified fp16 and Apache-2.0, with attribution to the
Laya authors and the original exporter recorded in the model card.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

squad-ci.yml shipped as an unfilled scaffold: it echoed "No build commands
configured" and exited 0 in four seconds, so it reported a passing status
check while testing nothing. It fired on PRs to main, which meant every PR
showed a green check that carried no signal.

It now runs the same restore/build/test the CI Build workflow runs,
including -p:NoBuild=false, which the BitNet native projects need or they
are silently skipped.

main is dropped from its triggers because ci.yml already covers main and
running the suite twice on one commit buys nothing. dev, preview and
insider do not exist yet, so this workflow is dormant until the dev-first
branches are created — at which point they get real CI instead of none.

Validated with actionlint.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

Both described the weights as simply coming from HuggingFace without
saying which repository, so neither reflected the move to the
owner-controlled mirror.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

…veats

The choice:11+ temperature clamp was tracked with an internal, model-wide flag
that callers could not read. A checkpoint can mix valid and invalid buckets, so
the actionable question is whether a given answer's own bucket was clamped.

Track clamping per temperature entry and surface it as CalibrationClamped on
ChoiceResult, ScoreResult and ProbabilityResult.

Running this against the English checkpoint shows the clamp is a guardrail, not
a fix: a 12-option choice still reports 100.0% confidence, because clamping to
0.5 bounds the tenfold sharpening but any temperature below 1.0 still sharpens.
The docs previously claimed clamping "keeps the number honest", which is wrong.
Correct that, and state plainly that ranking is reliable while the magnitude is
not.

Also document fp16 batch variance, measured at 1.2e-05 to 1.6e-05 between a
question asked alone and the same question asked inside a batch of four.

Add a DecisionCalibration sample that demonstrates both effects with live
numbers, and register it in the solution.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

The Pack step packs ElBruno.LocalLLMs.Decisions with --no-build, but the Build
step never built it and nothing else references it, so the Release output never
existed. Reproduced against a clean output directory: pack fails with NU5026,
'ElBruno.LocalLLMs.Decisions.dll to be packed was not found on disk'.

Any release published from this workflow would have failed at the Pack step.

Build the project alongside the other packable ones and run its test project in
the Unit Tests step so the package cannot ship untested.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968

…lish

Promote the live Unreleased section to 0.22.0, matching the existing dated
heading convention.

Also run ElBruno.LocalLLMs.BlazorComponents.Tests in the publish workflow. The
package was packed and pushed without its 23 tests ever running in the release
gate. Verified passing under the same Release configuration the gate uses.

Two stale Unreleased sections remain further down the file, describing features
that shipped long ago. They predate this work and are left untouched.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: c71e4170-a73a-4ac2-8c59-4b80834e7968