docs: document opt-in trial stability summaries - #907

Problem

Repeated trials can have the same mean while behaving very differently across attempts. The draft comparison contract adds an opt-in include_trial_stability field, but the public reference and CI workflow did not explain its denominator, eligibility rules, or limits.

Change

• Document the REST request flag and the conditional trial_stability response.
• Describe expected, recorded, missing, assessed, mean/min/max, and population standard deviation fields.
• Document paired same-trial candidate-minus-baseline deltas.
• Add an equal-mean example: baseline scores 0, 1 versus candidate 0.5, 0.5.
• Clarify that failed, invalid, duplicate, and unassessed evidence is excluded from score statistics; incomplete denominators remain explicit.
• Update the existing release-gate CI tutorial and external workflow guide. Trial stability remains descriptive and does not alter the gate or claim significance/reliability.

Validation

• node scripts/audit-links.mjs: 0 broken navigation/content links; 29 pre-existing orphan pages.
• npm run build: 1,190 pages built; Pagefind indexed 725 pages.
• git diff --check passed.
• Helper-backed check verified the equal-mean, paired-delta, incomplete-count, and REST flag-preservation examples.

Base: docs #904 (docs/TH-8579-case-release-gate).

Limitations: the option requires a compatible test deployment. This docs change does not establish hosted API, generated SDK, model, Observe, browser, or deployment qualification. No PR checks were reported at creation time.

Linear: TH-8586.