Add relevance gates, per-call size caps, and byte/cost accounting for tool outputs entering provider payloads

[baron]

Impact

Unbounded tool result bodies can flow into provider payloads without clear byte/cost visibility. Tool storms remain hard to see until context quality or reliability degrades. Sub-session model use, including intentional high-tier consultation, also needs cost attribution rather than prohibition.

Actual behavior

• Tool outputs can enter transcripts/payloads at full size without per-call caps or relevance gates before provider injection.
• There is no sufficiently visible accounting of bytes contributed per tool call, turn, session, or sub-session model call.

Expected behavior

• Configurable per-tool-call size cap with truncation markers and a retrieval path for full artifacts.
• Byte accounting per tool call, turn, and session, surfaced in diagnostics and/or usage UI.
• Sub-session model consultations, including high-tier models when intentionally used, are visible and attributed to the parent/session cost model.

Publishable evidence

• About 247 MB of rollout function-call output.
• Worst injections up to about 1.75M characters.
• About 11k exec_command calls observed in rollout data.

Scope / non-goals

In scope:

• per-call caps,
• truncation/retrieval UX,
• byte accounting,
• usage attribution for sub-session calls.

Non-goal:

• forbidding or discouraging high-tier model consultation by subagents. This issue is about visibility and attribution, not restriction.

Likely affected areas

• Sources/RepoPrompt/Features/AgentMode/Runtime/ToolTracking/
• Sources/RepoPrompt/Features/AgentMode/Runtime/Usage/
• Sources/RepoPrompt/Features/AgentMode/Runtime/Usage/Claude/
• Sources/RepoPrompt/Features/AgentMode/Runtime/Usage/Codex/
• Sources/RepoPrompt/Infrastructure/MCP/
• Sources/RepoPrompt/Features/AgentMode/Views/
• Tests/RepoPromptTests/AgentMode/
• Tests/RepoPromptTests/MCP/

Acceptance criteria

• Per-tool-call output cap exists with configurable limit.
• Truncated outputs include a clear marker and retrieval hint for the full artifact.
• Per-turn and per-session byte accounting is available in diagnostics or usage UI.
• Sub-session model usage is attributed to the parent/session in usage reporting.
• Tests cover cap behavior and accounting totals.

Interlinks

Part of #406. Related: #398, #399, and #401.