Make Codex-assisted Insights explain usage and recommend next actions instead of restating metrics

[thrr87]

What happened

Running Analyze metadata consumed Codex allowance but returned “You have 37% usage remaining,” duplicating the current-state header, Graphs, and Facts. It did not explain the usage pattern, connect measurements, identify what deserves attention, or suggest an improvement.

The same limitation applies to Source Content analysis: a successful result can still amount to a selected fact rather than an interpretation.

What I expected

Codex-assisted Insights should be the interpretation and decision layer:

• Synthesize the relevant available measurements instead of repeating one.
• State the important pattern or implication for the selected range and why it matters.
• Offer a concrete, testable next action when the evidence supports one.
• Separate measured evidence, interpretation, recommendation, and uncertainty.
• Withhold the Insight and name missing evidence when no useful conclusion is supported.

Deeper analysis must remain bounded: use only authorized evidence, respect the selected scope and filters, retain provenance/Coverage/Confidence/freshness, and avoid unsupported causal or cost claims.

Steps to reproduce

1. Open Codex Limits with fresh weekly usage data and the supported Codex analysis profile available.
2. Open Insights.
3. Click Analyze metadata.
4. Wait for the Codex-assisted result.
5. Compare the result with the current-state header, Graphs, and Facts.

Additional context

A successful result should not be semantically equivalent to a single value already visible elsewhere. When multiple reliable fields are available, it should relate them into one useful finding and, when supported, at least one actionable recommendation. Insufficient evidence should produce a clear withheld state rather than a fact restatement.

For Source Content analysis, findings and recommendations must be backed by validated references without quoting or reproducing the content. Existing consent, privacy, model-profile, isolation, provenance, and Analytics Overhead behavior must remain intact.

[thrr87]

Implemented in #64.

Codex-assisted analysis now returns a bounded structured finding, why it matters, a conditional next action, and references to at least two validated evidence fields. Weak evidence produces a clear Not enough evidence result instead of a repeated metric. Source Content uses the same contract and rejects invalid references, unsupported claims, and verbatim content.

A real isolated metadata request related account activity, local activity, and usage per token, then proposed a matched-interval comparison. The separate QA window rendered the deeper result successfully.

[thrr87]

Follow-up verification for the sparse-history case reported during local QA:

• Metadata analysis now receives the pace forecast already used by the header, including current versus recommended pace, expected headroom, Coverage, and Confidence.
• Low-coverage guidance can produce a cautious, actionable Insight instead of “Not enough evidence.”
• When neither pace guidance nor two reliable fields are available, the app prevents the Codex request locally so it does not spend allowance on a predictable withheld result.
• A real isolated request returned “Room to increase usage,” a why-it-matters explanation, and a conditional next step. The separate QA window kept the result visible after Refresh.

Verified by 564 passing Swift tests and a signed production build in #64.