Added JSON protection and selective text compression based on tenant schema

[PredictabilityAtScale]

What changed

• Protect every valid JSON object and array from probabilistic model compression; TOON is now an optimization gate rather than a protection gate.
• Add tenant-authorized <compress-json policy="..."> blocks for bounded compression of explicitly allowlisted narrative string paths.
• Add a project changelog, a rendered /changelog page, navigation links, and tagged-JSON documentation.
• Bump the deployment version and include the changelog in the Cloud Run image.

Why

The benchmark exposed a fidelity risk: model compression could preserve numeric values while dropping their structured field labels. The new JSON pipeline keeps structure deterministic and confines model compression to opt-in tenant-approved string leaves.

Validation

• python -m pytest -q --basetemp=.pytest-tmp/changelog-full (155 passed)
• python -m ruff check app tests
• git diff --check

Operational note

model_force can make one call per accepted tagged JSON value, up to json_value_max_values, plus the outer prompt call.