Qualify autonomy per workflow

[yihanzhu]

Problem

ystack says autonomy is earned in stages, but it does not yet say what exactly earns authority or what a stage must return. That can let trust drift from one task, model, or environment to another, and it can make a useful no-code investigation look like failure.

Outcome

Make one bounded roadmap amendment:

• authority belongs to a recorded workflow scope, never to an agent in general;
• each execution environment qualifies separately;
• policy records which changes invalidate qualification and which gate runs again;
• each stage attempt has an explicit status, and executed work has an evidence-backed result;
• verification evidence stays typed and bound to the relevant identities;
• portable core carries result, evidence, and qualification references, while later initiatives own policy and enforcement;
• the first shadow workflow is narrow and read-only when practical, with incident reproduction as the preferred case.

Constraints

• Do not enable autonomous writes.
• Do not change the human-merge or read-only-reviewer rails.
• Keep the core harness-, model-, forge-, and CI-neutral.
• Do not add the separate verification-pack, eval-protocol, skill-lifecycle, metrics, model-policy, or audience-interface work to this PR.
• Change ROADMAP.md and only the README rollout sentence needed to keep the
  public summary in sync.