fix: remove the default defect-group cap

[svozza]

A review can have fewer than ten finding entries and still be rejected by the shipped four-defect-group cap. On Cal.com, that cap pushed confirmed concerns into residual risk and repeatedly missed the Office365 destination defect.

Ship review.max_distinct_groups: null and state the effective group policy in the generated constraints. Keep the ten-entry maximum, required group field, numeric group range and all provenance/output checks. Numeric consumer caps remain supported and are disclosed in the prompt. Runtime files match the evaluated no-cap candidate exactly; the additional changes update the default-policy tests and architecture/ADR documentation. An evaluation-only correction also addresses the grouping fixture’s documented rejection of an optional root-cause finding; its exact path/line/diagnosis matcher preserves required callers and rejects unrelated or split groups. No reviewer input changes accompany that correction.

The completed direct comparison against shipped 8b91cb5 used three paired repetitions across six frozen source snapshots (36 reviews). Known target delivery improved from 18/24 to 21/24, recovering Office365 in all three candidate runs versus none in control. Summed review time fell 5.57% and submission rejections fell 5 to 0. Both arms still had one false structured finding and incomplete repair advice; repeated nonexhaustive targets do not establish general accuracy. Raw source assessments remain private/local.

Validation:

• Full deterministic suite: 2,400 passed, 2 fixture-dependent skips.
• Type checking and diff checks passed.
• Boundary tests accept ten distinct groups, reject eleven entries, and preserve configured-cap enforcement.
• Initial CI: two raw passes, then a false grouping failure documented by the fixture itself. All three runs remain retained. Uniform regrading corrects the false failure and exposes an incorrectly split root in the second run.
• Final-head CI is green: deterministic tests/type checking and three fresh full evaluation passes, totaling 126/126 review scenarios and 18/18 plan scenarios, with zero API errors. Earlier runs remain diagnostic evidence and do not count toward this gate.

Final-head evaluation runs: 1, 2, 3. Deterministic CI: Quality check.