fix(eval): preserve budget termination outcomes

[Sayt-0]

Summary

• preserve native budget_exceeded outcomes as allow-listed evaluation termination metadata
• keep the chronological termination marker and exactly one associated assistant stop message across transcripts, SQLite, and sessions JSON
• expose optional eval_result.termination data without changing scoring or error semantics
• keep message_added payloads private by carrying the budget stop through a dedicated narrow runtime DTO
• document the result contract and backward-compatible interpretation

Closes #3845

Acceptance criteria

Expectation	Implementation
Native event parsing	budget_exceeded maps to reason, budget, limit, used, max, config_path, and message only
Safe metadata	copied strings are UTF-8 normalized, control-filtered, bounded, and omitted when unusable
Chronology	the termination item is emitted at the event position, immediately followed by its assistant stop message
Duplicate prevention	repeated equivalent budget events do not duplicate the marker or stop message
Narrow message handling	all message_added events remain payload-free and ignored by evaluation reconstruction; only the dedicated budget stop_message DTO is retained
SQLite persistence	termination items reuse the existing session_items.message_json storage with item_type = termination, requiring no migration
Sessions JSON	termination marker, assistant stop message, and optional eval_result.termination survive serialization and reload
Result semantics	termination remains informational and does not alter passed, failures, errors, or scoring inputs
Compatibility	old sessions and results without termination data continue to load unchanged
Isolation	normal, error, ordinary stream-stop, and foreign-session events do not receive or disturb budget termination data

Validation

• env -u HTTP_PROXY -u HTTPS_PROXY -u ALL_PROXY -u http_proxy -u https_proxy -u all_proxy go test ./...
• go build ./...
• go vet ./pkg/runtime ./pkg/evaluation ./pkg/session
• golangci-lint run ./pkg/runtime/... ./pkg/evaluation/... ./pkg/session/...
• go run ./lint .
• npx --yes markdownlint-cli2@0.22.1 configuration/budget/index.md from docs/
• ./scripts/docs-check-canonical.sh
• git diff --check

Manual validation

Not run. A live budget-triggered evaluation would require a configured model/provider. Deterministic runtime, parser, chronology, persistence, compatibility, isolation, and security coverage is included in the automated tests.