feat(handoff): add host-advisory rollover triggers for long sessions #1787

Feature description

Add host-advisory rollover trigger observations for long sessions.

PowerContext already has Handoff as the work-continuity boundary, and PR #1783 proposes/implements advisory rollover_reasons for Handoff preparation. This issue tracks the next step: let host integrations report bounded observations that can suggest or prepare a Rollover Handoff when a session is becoming a poor carrier of current work state.

Codex prior art to review:

• [codex] Add new context window tool openai/codex#27488 - [codex] Add new context window tool
• Allow models to enable token budgeting by default openai/codex#41803 - Allow models to enable token budgeting by default
• Add experimental context management activation openai/codex#42385 - Add experimental context management activation

The portable idea is not the provider-specific ability to open a new context window. The portable idea is the decision boundary: when the active session is long, budget-constrained, compacted, or low-quality, preserve the work state outside the old context before continuing elsewhere.

Problem and proposed solution

Long agent sessions can accumulate memory rot before they hard-fail on token limits. Hosts may have signals such as:

• approximate token/context budget pressure;
• compaction count or imminent compaction;
• session age or number of tool turns;
• model-reported context quality problems;
• user-requested fresh-session continuation.

PowerContext should accept these as advisory host observations and translate them into a bounded Handoff preparation path, not into agent control.

Proposed behavior:

• Define a host observation shape for context budget, compaction, and context-quality signals.
• Allow host integrations to attach those observations as Source evidence where policy permits.
• Map observations to existing Handoff rollover_reasons such as host_context_budget, host_compaction, and context_quality.
• Support manual/advisory/automatic-draft levels, while keeping automatic commit out of scope.
• Keep the flow fail-open: if observations are missing or unsupported, normal Handoff and PreparedContext behavior continues.

Alternatives considered

• Let each host implement rollover heuristics privately. This is easy, but makes behavior hard to explain and impossible to evaluate across hosts.
• Store token-budget observations in Memory. That would pollute durable project memory with session-local telemetry.
• Let PowerContext open or manage agent sessions. That crosses the product boundary; PC should preserve state, not run the agent.

Additional context

Related PowerContext work:

• PR RFC: Handoff session rollover #1783: Handoff session rollover RFC and initial rollover_reasons implementation.
• tracking(continuity): 宿主历史导入与可评估的上下文连续性 #1574: host history import and evaluable context continuity.
• Tracking Issue: make PowerContext work continuity reliable, explainable, and cross-host #1352: reliable, explainable, cross-host work continuity.

Acceptance criteria:

• A host can report context-budget/compaction observations without changing committed Handoff content identity.
• The observations can drive Handoff prepare with the appropriate rollover_reasons.
• Unsupported hosts remain compatible and do not need to provide these signals.
• The docs clearly state that observations are untrusted unless backed by authorized Source evidence.