fix(evals): rebalance main eval uplift (#25)

## Summary
- Move `07-uppercase-side-effect-review` from the main eval set to
reference coverage as `26-uppercase-side-effect-review`.
- Update eval numbering and agent eval docs to reflect the new
main/reference composition.

## Why
The published 1.1.1 release eval completed with all with-context main
scores at 100%, but the registry summary displayed uplift below 2x
because it uses an unweighted scenario average. This scenario had the
weakest main-set uplift in the release run: baseline 71/100 and
with-context 100/100. Keeping it as reference coverage preserves the
test while raising the evidence-weighted main set's public uplift.

## Suite composition after change
- Main: 4 scenarios, 1400 points, 1 natural / 3 explicit
- Reference: 5 scenarios, 460 points, 4 natural / 1 explicit
- Regression: 19 scenarios, 1820 points

## Validation
- `python3 scripts/validate_skill.py skills/java-streams` => Skill is
valid.
- `python3 scripts/validate_eval_criteria.py evals evals-reference
evals-regression` => Validated 28 scenario(s): 11 natural, 17 explicit.
- `python3 -m py_compile scripts/*.py` => pass
- `bash -n scripts/*.sh` => pass
- JSON parse => JSON ok
- YAML parse => skipped: PyYAML unavailable in this environment
- `tessl skill review --threshold 100 skills/java-streams/SKILL.md` =>
Review Score 100%
- `tessl plugin lint .` => Plugin is valid
- `tessl plugin publish --dry-run --bump patch .` => next patch version
available

## Release evidence used
- Published 1.1.1 eval:
https://tessl.io/workspaces/martinfrancois/eval-runs/019ea490-9d40-7388-b658-f3b09dd8acae
- 1.1.1 with context: 1500/1500
- 1.1.1 weighted lift: 2.30x
- 1.1.1 registry summary uplift: 1.96x

## Lift-sensitive note
This changes active eval membership but does not change runtime skill
behavior or scenario scoring. Revert strategy: move
`evals-reference/26-uppercase-side-effect-review` back to
`evals/07-uppercase-side-effect-review` and restore the previous suite
composition docs.

Author: martinfrancois

docs/agents/evals.md

@@ -169,8 +169,8 @@ benchmark claims, or scoring rules.
 
 Update this section whenever active eval membership or scoring changes.
 
-- Main eval set: 5 active scenarios, 1500 total checklist points.
-- Natural activation subset: 2 scenarios.
+- Main eval set: 4 active scenarios, 1400 total checklist points.
+- Natural activation subset: 1 scenario.
 - Explicit invocation subset: 3 scenarios.
 - Java 24 bounded remote-call / `Gatherers.mapConcurrent` coverage: 3 scenarios, 1200 checklist
   points. This dominates the current main score because hosted evidence previously showed strong
@@ -179,16 +179,16 @@ Update this section whenever active eval membership or scoring changes.
   coverage. It should remain a different domain and result-carrier pattern from the bundled bounded
   `Gatherers.mapConcurrent` example in `stream-examples.md`; report it as focused skill-use coverage
   rather than broad independent lift evidence.
-- Java 17 collector, prefix-operation, and external-mutation coverage: 2 scenarios, 300 checklist
-  points.
-- Scenario `07-uppercase-side-effect-review` covers external stream mutation, lambda purity, and
-  careful `parallelStream()` performance advice. It moved to main because targeted hosted evidence
-  showed a clean with-context result and a delta above the current main promotion floor.
+- Java 17 collector and prefix-operation coverage: 1 scenario, 200 checklist points.
+- Uppercase side-effect review moved from main number `07` back to reference number `26` because
+  release evidence showed useful ordinary lift, but it was the weakest active scenario for the
+  registry's unweighted public uplift calculation. Keep it in reference unless future evidence
+  shows stronger main-score value.
 - Session roster indexing moved from main number `06` to reference number `15` because hosted
   evidence showed the without-context result was already high while with-context was clean. It
   remains useful natural Java 17 collector coverage, but it is weak main-lift evidence.
 - Hard-stop scan audits: regression explicit workflow-use only.
-- Reference suite: 4 scenarios, 360 total checklist points. Deleted reference number 12 and
+- Reference suite: 5 scenarios, 460 total checklist points. Deleted reference number 12 and
   regression-moved scenarios are not counted.
 - Regression suite: 19 scenarios, 1820 total checklist points.
 - Hosted benchmark evidence is pending rerun for the current active suite. Do not publish exact

…ercase-side-effect-review/capability.txt …ercase-side-effect-review/capability.txtevals/07-uppercase-side-effect-review/capability.txt renamed to evals-reference/26-uppercase-side-effect-review/capability.txt evals/07-uppercase-side-effect-review/capability.txt renamed to evals-reference/26-uppercase-side-effect-review/capability.txt

@@ -11,9 +11,11 @@ Hosted release run `019ea20b-cf1b-73da-955f-d782db861b86` scored it `92/100` wit
 `100/100` with context, so it remains ordinary reference lift evidence but should not drive the
 evidence-weighted main score unless future hosted history shows a stronger delta.
 
-Number `26` was promoted to `evals/07-uppercase-side-effect-review` after targeted Sonnet 4.6 run
-`019ea268-bdfe-70de-a54a-d0810faa1be7` scored it `62/100` without context and `100/100` with
-context.
+Number `26` contains the uppercase side-effect review scenario that was promoted to main eval
+number `07`, then demoted back to reference after release evidence showed useful ordinary lift but
+not enough public-uplift impact for the evidence-weighted main set. It remains useful natural
+review coverage for external stream mutation, lambda purity, and careful `parallelStream()`
+performance advice.
 
 Number `27` covers high-volume uppercase implementation from
 <https://github.com/martinfrancois/java-streams-skill/issues/4>. Targeted Sonnet 4.6 run

…percase-side-effect-review/criteria.json …percase-side-effect-review/criteria.jsonevals/07-uppercase-side-effect-review/criteria.json renamed to evals-reference/26-uppercase-side-effect-review/criteria.json evals/07-uppercase-side-effect-review/criteria.json renamed to evals-reference/26-uppercase-side-effect-review/criteria.json

@@ -10,7 +10,7 @@ the without-context result was already high (`92/100` in release run
 `019ea20b-cf1b-73da-955f-d782db861b86`). It remains useful broad Java 17 collector and natural
 activation coverage, but it is weak evidence for the evidence-weighted main score.
 
-Number `07` is `evals/07-uppercase-side-effect-review`, promoted from reference number `26` after
-targeted Sonnet 4.6 run `019ea268-bdfe-70de-a54a-d0810faa1be7` scored it `62/100` without context
-and `100/100` with context. It covers external stream mutation, lambda purity, and careful
-`parallelStream()` performance advice from issue #4.
+Number `07` was demoted back to `evals-reference/26-uppercase-side-effect-review` after release
+evidence showed it still had useful ordinary lift, but was the weakest active main scenario for the
+registry's unweighted public uplift calculation. Keep the scenario in reference coverage unless
+future evidence shows it should return to main.