Make default task context concise and evidence-complete

[grove]

Default task context could show constraint IDs and assessments before their actual source rules, and could blur static reads, future checks, and fallback confidence. The human output now puts the action and readiness first, followed by complete constraint statements, scope/lifecycle qualifiers, exceptions, rejected alternatives and conflict counterparts. Constraint facts appear once; local observations, imported reports, hypotheses and checks retain distinct labels.

Fallback guidance asks for an evidence or compatibility check before choosing an implementation. Reports and rejected history no longer become “retained rules,” and bounded retrieval-only output explicitly says “Decision not assessed.” The existing schema-5 JSON contract, evidence IDs, source snapshots, privacy boundaries and complete-envelope budget checks remain in place.

Validation:

• Eight human-output goldens, including missing evidence, denied inspection, stale reports and budget insufficiency.
• Default 3,000-token JSON/human CLI check and complete evidence-group checks under smaller budgets.
• cargo test --all-targets --locked: 438 passed, 2 ignored.
• cargo fmt --all -- --check, strict Clippy, and release build passed.
• Python 3.14 evaluation suite: 271 tests, 1 skipped; two process-double failures in the sandbox passed when both affected suites were rerun on the host (14 tests).

Live providers and cross-platform CI were not exercised locally. These fixture checks do not establish recommendation quality or productivity gains.

Fixes #84