Add a TokenSave repository-context plugin

## Parent and milestone

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

## Problem

Agents repeatedly scan repositories with grep and file reads. TokenSave offers a local semantic graph, MCP tools, branch-aware indexes, and cross-agent support, but must prove correctness and operational fit in Durex.

## Outcome

Add TokenSave as an optional explicit context provider only after it passes the benchmark gate, with health, freshness, branch, privacy, and fallback behavior visible to operators.

## Scope

- Integrate through the standard context-provider/MCP boundary.
- Validate index presence, freshness, branch mapping, supported language, and resource bounds.
- Record context provenance and token estimates.
- Fall back to normal agent exploration when configured policy permits.

## Non-goals

- Making TokenSave mandatory.
- Enabling a second semantic graph concurrently by default.
- Trusting savings-ledger estimates as Durex benchmark evidence.

## Acceptance criteria

- [ ] Stale, missing, corrupt, unsupported-language, and branch-switch cases are tested.
- [ ] Disabling the plugin restores baseline behavior without changing task data.
- [ ] No repository content leaves the host unless explicitly configured and documented.
- [ ] Targeted, failure-path, and conformance tests pass.
- [ ] Architecture, plugin-developer, configuration, and operator documentation is updated.

## Validation

- Run benchmark-selected fixtures with and without TokenSave at a pinned version.

## Dependencies

#34 and #36; implementation starts only if the benchmark recommendation is positive.

Start at Durex's standard context-provider/MCP boundary and review dependencies #14, #15, #34, and #36 before implementation. Run the benchmark-selected fixtures at a pinned TokenSave version with and without the plugin. Done means the acceptance cases pass, fallback and privacy behavior are verified, and the listed architecture, plugin-developer, configuration, and operator documentation is updated.

Written by the indexing model from the issue text.