feat(usage): track native compaction requests

[okbexx]

Problem

Native Remote Compaction V2 uses the normal POST /v1/responses route, so usage logs currently preserve only the transport shape (stream / WebSocket) and lose the semantic fact that a request was a remote compaction. Operators cannot identify or filter confirmed compaction rows without unsafe inference.

Approach

• add a dedicated usage_logs.native_compaction_v2 boolean instead of overloading request_type
• capture the existing protocol-positive marker before asynchronous usage recording
• preserve the existing transport dimensions (request_type, stream, openai_ws_mode)
• propagate the flag through single/batch/best-effort inserts, reads, DTOs, filters, query caches, trend/model/group/breakdown queries, and raw-query fallback selection
• expose a positive-only Compaction filter and an additional badge in both admin and user usage views
• keep cyber policy precedence: a blocked compaction request remains request_type=cyber while retaining the semantic flag

Safety and compatibility

• classification remains limited to /v1/responses + stream=true + input[].type=compaction_trigger
• no request body, compaction content, credentials, or authorization data is persisted
• historical rows are not backfilled or claimed as protocol-confirmed compactions
• Channel Monitor V2 continues counting successful compactions because transport request types are unchanged
• migration defaults the new non-null column to false; running it twice is idempotent

Verification

• Go 1.27 targeted packages:
  • ./internal/repository
  • ./internal/handler
  • ./internal/handler/admin
  • ./internal/handler/dto
• Go 1.27 focused service regressions:
  • marker persistence
  • stream/compaction orthogonality
  • cyber request-type precedence
• PostgreSQL 17 disposable migration validation:
  • default false
  • explicit true
  • NOT NULL
  • idempotent re-run
• frontend:
  • npm run typecheck
  • 4 focused Vitest specs for filters, table badges, and admin/user query propagation
  • ESLint on all 14 changed frontend source/test files
• git diff --check

All Docker-based validation used --rm containers with bounded temporary caches; no persistent Go cache volume was created.