feat(context): govern and enforce Agent context budgets

[JasonW404]

Summary

Implements the complete context-budget P0-P3 plan: safe capacity configuration, governed model profiles, final Provider-request budget enforcement, compaction, and observable health/usage workflows.

Backend and runtime

• Validate and persist explicit model capacity limits without manufacturing unknown values.
• Add model identity, capability-profile matching, field-level provenance, tokenizer governance, count probing, catalog adoption, and lifecycle health.
• Meter the exact final Provider request, enforce soft/hard budgets, compact context safely, support one-shot recovery, and calibrate estimates from Provider usage.
• Persist content-free request evidence and expose tenant-isolated operational metrics.
• Add health, catalog, remediation, and monitoring APIs with existing authorization boundaries.

UI and UX

• Add validated capacity fields and governed suggestions to model create/edit flows.
• Add model-capacity health, lifecycle, review, and field-level remediation UI.
• Show live and persisted context usage, token-component breakdowns, thresholds, count source, compaction savings, and optimization reasons in chat.
• Add context-budget operations metrics and English/Chinese copy.

Other changes

• Add governance and observability database migrations.
• Add acceptance-driven backend, SDK, migration, monitoring, and frontend regression coverage.
• Preserve old model rows, monitoring records, and conversation history.

Verification

• All P0-P3 acceptance criteria pass in the specification evidence.
• Final affected backend gate: 100 passed.
• Frontend type-check and production build: PASS.
• Final images: nexent/nexent:context-budget-p3-a33295a and nexent/nexent-web:context-budget-p3-a33295a.
• Full browser acceptance passed with permanent tenant and real Qwen3.7-plus.
• Normal trace: f732eb0e01f72174b1234b1ea55e4463.
• Compaction trace: 9280aa7369d7ea56e546bef742a3fdae; 29,603 to 21,623 tokens, saving 7,980 (27%).
• Tenant ADMIN remediation and ordinary USER permission checks passed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.