product: calculate and surface per-session evidence health

[Siddhant-K-code]

Goal

Tell a reviewer whether a session record is complete enough for the decision at hand, and explain why it may be partial or ambiguous.

This implements the evidence-trust requirement from the Verio telemetry learning: a clean timeline must not be presented as proof of complete capture.

Problem

The current UI can show a running or completed session and raw events, but it does not calculate a first-class evidence-health result. Missing boundaries, unpaired events, provider limitations, stale metadata, or export failures can be easy to miss.

A generic percentage would be misleading because AgentTrace cannot count events a provider never exposes. Health must combine observed structural checks with the declared limits of the capture method.

Scope

Define a deterministic, versioned evidence-health result with:

• status such as healthy, partial, unknown, or invalid;
• observed session start and end boundaries;
• capture provider, method, and applicable capture-matrix version;
• unpaired tool calls/results and other broken event relationships;
• timestamp regressions or suspicious discontinuities;
• stale or unfinalized session metadata;
• recorded export failures or dropped-event signals when observable;
• explicit provider blind spots; and
• machine-readable reason codes plus human explanations.

Expose the same result in:

• a CLI inspection path;
• replay/session summary output;
• JSON or MCP/API session metadata; and
• the local web dashboard session header.

A status must never imply that unobservable provider activity was complete.

Design constraints

• Keep the calculation local and dependency-free.
• Separate observed defects from declared provider limitations.
• Preserve source provenance for every reason.
• Do not infer misconduct, policy failure, or compliance from missing evidence.
• Make the rules testable against complete, deliberately partial, legacy, and malformed fixtures.

Acceptance criteria

• The health schema and reason codes are documented and versioned.
• Missing start/end boundaries, unpaired tool events, timestamp regressions, and stale metadata are detected.
• Provider blind spots are sourced from the capture matrix rather than guessed.
• Complete, partial, legacy, active, and malformed session fixtures have deterministic results.
• CLI, JSON/API, replay, and local dashboard show consistent status and reasons.
• No health state claims capture completeness beyond what the provider exposes.
• Export/drop failures are shown when AgentTrace has observable evidence of them.