Evaluate safe prompt compression for derived prose context

## Parent and milestone

- Parent issue: #24
- Milestone: not scheduled; implementation depends on #14 and starts after #15.

## Problem

Caveman and LLMLingua can reduce prose tokens, but compression may alter constraints, identifiers, negation, or acceptance criteria in long-running coding tasks.

## Outcome

Evaluate compression only for derived summaries and retrieved prose, with protected spans, semantic checks, provenance, and original-content fallback.

## Scope

- Compare deterministic terse summarization, Caveman-style transformation, and LLMLingua on derived prose fixtures.
- Protect code, commands, paths, identifiers, numbers, negation, requirements, security rules, and user-authored prompts.
- Measure token reduction, latency, semantic preservation, and downstream task correctness.
- Define a safe allowlist of content classes or reject compression entirely.

## Non-goals

- Compressing source code, patches, approvals, original user prompts, exact errors, or final documentation.
- Using compression claims without Durex-specific correctness tests.
- Enabling compression by default before review.

## Acceptance criteria

- [ ] Protected spans remain byte-identical.
- [ ] Semantic regression fixtures cover negation, numeric bounds, paths, identifiers, and security constraints.
- [ ] Original prose remains recoverable and linked from the transformed envelope.
- [ ] Targeted, failure-path, and conformance tests pass.
- [ ] Architecture, plugin-developer, configuration, and operator documentation is updated.

## Validation

- Run repeated downstream task fixtures and reject any mode that lowers completion correctness beyond the approved threshold.

## Dependencies

#33 and #36; intentionally experimental until correctness evidence exists.

No implementation files or tests are named. Start by reviewing parent issue #24, dependencies #14, #15, #33, and #36, then identify the existing context and validation entry points. Done means protected spans and recoverability are demonstrated, correctness thresholds pass, and the listed documentation is updated.

Written by the indexing model from the issue text.