[P1][EVAL-03] Outcome benchmarks and anti-wrapper tests

Programme: #1
Epic: #33

## Objective
Measure whether the product actually delivers a persistent user-owned computing experience that outperforms ordinary browsing/chat assistance on verified outcomes—not merely a better-looking agent demo.

## Scope
- Scenario suites for research, shopping comparison, synthesized reading, social, coding, files, browser automation and native desktop tasks.
- Compare ordinary browsing, chat-based assistance using the same underlying tools/models, and persistent workspace mode.
- Anti-wrapper tests: close chat, restart app, disable inference, change provider/model and continue using the workspace.
- Repeated trials for nondeterministic tasks with confidence intervals and failure taxonomy.
- Human review for relevance, usability, evidence quality and corrective intervention burden.
- Metrics: verified completion, elapsed time, interventions, forced Original-mode handoffs, source/evidence quality, user understanding of authority, latency/cost/energy.
- Held-out private scenarios plus staged low-risk production canaries.

## Evaluation rules
- The model cannot be the sole grader of its own completion.
- Provider/model comparisons use pinned versions/configurations and disclose them.
- Do not use time-spent-in-product as the north-star metric; forced source inspection and healthy provenance use must not be penalized.

## Acceptance criteria
- [ ] Workspace remains directly usable after chat closure/restart with inference disabled for supported deterministic interactions.
- [ ] Provider switch preserves GoalContract, workspace, evidence and artifacts for declared scenarios.
- [ ] Benchmarks report verified completion separately from attempted/self-reported completion.
- [ ] Nondeterministic results include repeated trials, confidence intervals, interventions and failure classes.
- [ ] Comparisons use the same underlying tools/models where the goal is to isolate the interaction/runtime advantage.
- [ ] Release dashboards expose cost, latency, energy and correctness together rather than optimizing a single benchmark score.

## Dependencies
- EVAL-01
- WS-01
- WS-02

**First phase:** P1
**Maturity target:** P7 (continuous)
**Owner:** security-evaluation-release

Start by reading the EVAL-01, WS-01 and WS-02 dependencies and defining which declared scenarios they support. Establish the benchmark and human-review entry points for deterministic and nondeterministic trials. Done means the acceptance criteria are measured separately, with repeated-trial confidence intervals, failure classes, provenance checks, and combined cost, latency, energy and correctness reporting.

Written by the indexing model from the issue text.