Prepare v1.2.0: voice calibration and claim-preserving edits

[milock]

Problem and result

Humanizer already has basic voice calibration, but its generic rules can override approved voice choices and its worked examples introduce facts absent from the input. This prepares v1.2.0 with evidence-based calibration and claim-preserving editing, adapted from Michael Lock's September deslop work.

• Select the publisher's voice source, capture six observations, and compare the result against them. Explicit preferences override generic rules; calibration stays transient unless saving is requested.
• Preserve facts, uncertainty, negation, rankings, and asks. Flag missing evidence and hollow drafts without inventing examples or metrics.
• Keep existing output headers, remove unsupported detector-performance claims from active guidance, and replace contradictory examples.
• Include voice templates and documentation in installations; repeated installs do not nest resource folders.

Validation

Skill validator, shell syntax, whitespace checks, local Markdown links, and fresh/repeat installation checks passed. Independent review checked contracts and contradictions. A forward exercise preserved voice preferences, uncertainty, a date, a number, a request, a technical term, and code. Two qualitative baseline/current simulations found clearer instructions and no observed regressions; these are not controlled model benchmarks.

Release state

Draft preparation only. CHANGELOG marks 1.2.0 Unreleased; no tag or release has been published. See docs/release-1.2.0.md for validation limits and publication steps.