Mid-word tab-to-complete needs a labelled calibration set before its plausibility floor can be tuned

## What happens

#58 asks for suggestions that complete a word cut mid-token (`gi` → `git status`), and #338 is the open pull request for it. The judgement still rests on one number: `Verification.plausibilityFloor = -6.0` (`Sources/UttrflowPredict/Verification.swift:63`). With #338 applied, an invented continuation such as `gi` → `gizmo --frobnicate` also passes that floor. No labelled set of real command-line prefixes exists to say where the floor should be for mid-word cuts, so any value chosen now is a guess that trades wrong suggestions against missing ones without measuring either.

## Why it matters

A confident nonsense completion in a terminal is worse than no completion: accepting it by habit runs the wrong command.

## What is needed

A calibration set, committed as fixtures, of:

- real command-line prefixes cut mid-token (for example `gi`, `do`, `kub`, `swift b`), each with
- one or more sensible continuations that should be offered, and
- plausible-looking nonsense continuations that must be refused.

It should cover several tools and argument shapes, and contain no personal paths, hostnames or real history. Invented but realistic values only.

## Acceptance criteria

- The fixture set exists in the evaluation corpus, with at least a few dozen prefixes across several tools.
- A bake-off command (`uttrflow-bakeoff score` or a new mode) reports, for a given floor, how many sensible continuations are kept and how many nonsense ones pass.
- The floor used for mid-word completion in #338 is chosen from that report, and the report is recorded in `Docs/`.

`help wanted`: building a good set benefits from people who live in a terminal with tools other than the maintainer's.

Start by reading Sources/UttrflowPredict/Verification.swift at the plausibility floor and reviewing the behavior proposed in #338. Inspect the evaluation corpus and the existing bake-off entry point, if present, before defining fixtures for mid-token prefixes. Done means the corpus, floor-scoring command, selected floor, and recorded report in Docs satisfy the acceptance criteria.

Written by the indexing model from the issue text.