Operator-prose register per D-215/D-216: rewrite getting-started.md + realize the audit doc-probe register judgment

[StarshipSuperjam]

Design spec landed (engine-planning): decision-log.md D-215 (authorize) + D-216 (resolve); the sharpened law in glossary.md "Operator-communication law" and systems/surfaces/docs/README.md; the audit teeth in systems/guardrails/audits/README.md posture law 3.

Two parts.

1. Rewrite docs/getting-started.md to the operator-communication law. The shipped getting-started reads as condescending ("talking down to a child") — a conformance miss against the law's register edge: address the operator as the capable adult they are; never condescending, never a register pitched below them, never explaining at length what they plainly already grasp as if they could not. The substance edge (clear, complete, never dumbed down) still holds; register is the orthogonal axis the doc fails — a doc can be complete and accurate yet talk down.

2. Realize the audit doc-probe's register judgment in the audit persona's instructions. The cold-context random-target doc-probe must, when the pick is an operator-facing doc, judge register against the law's rubric:

• conservative and audience-aware — weigh the doc against its intended reader (a thorough beginner-oriented orientation doc is not condescension for being thorough); flag only clear talking-down, never stylistic preference;
• recommend-not-block — a register finding is surfaced at the audit tier (a recommendation the operator adjudicates), never a merge gate;
• a register finding the operator judges wrong is a legitimate decline, never a manufactured nomination (posture law 1's no-quota / declining-is-a-real-choice).

[StarshipSuperjam]

Partial update from the audit-library build (PR #156, Slice 1): the audit persona (.claude/agents/audit.md) now carries posture-law-3's register-judgment edge in its instructions — when its cold random-target pick is an operator-facing doc, it weighs the doc's register (addressing the operator as a capable adult vs. talking down/over-explaining), flagging only clear talking-down, judged against the doc's intended reader, with a wrong call a legitimate decline. That delivers the persona-side of this issue's "realize the audit doc-probe register judgment" (D-215/D-216).

Two parts of #135 stay open and are out of scope for #156:

• The register judgment only runs once the scheduled audit fires — that lands in a later audit-library slice (the audit-prep cron workflow), not this static-artifacts slice.
• The getting-started.md rewrite per D-215/D-216 is untouched and remains this issue's own scope.

[StarshipSuperjam]

Scope extended by the engine-planning recentering pass (D-225 → D-228).

The operator-communication law this issue realizes has been deepened. Beyond the register edge (D-215/D-216), decision-log.md D-225 reframes the operator from deficiency to role + burden-of-proof — "a capable adult who builds through the engine rather than by reading code; not assumed to read code, so the burden of proof is on the engine; the AI is what's being made trustworthy, not a fragile user" — and adds the no-banned-word-list rule (clarity over jargon is judgment, never a filter).

Two adjustments to this issue's scope:

1. Part 1 (getting-started.md rewrite): also remove the rescue tone / deficiency framing — e.g. "you do not need to understand the inner workings", "not the Engine turning against itself" — not only condescension. Conforming models: README.md, .engine/conduct/defaults.md.
2. Part 2 (audit doc-probe): per D-226/D-227 the doc-probe's register/jargon judgment now covers operator-facing tool strings, not docs alone — realize that in the audit persona (also tracked in the governance-scrub issue). The "sweep all operator copy" is a real cold-audit pass over the tool-rendered operator strings, not a quick confirm.

Refs: engine-planning decision-log.md D-225/D-226/D-227, risks.md R30, systems/guardrails/audits/README.md.