[dpuglielli]

Summary

Refs #185; builds on merged #184.

Adds an explicitly configured Claude display mode that injects observed token categories and elapsed wall time into context before prompts and after tool use. Disabled and collect modes remain silent; normal plugin installation registers no resource hooks. Codex supports passive collection only. Mantra is bumped to 0.9.0.

The eight-run comparison found seven passing outcomes: all four collect runs passed; the display clarification run missed a necessary question and failed. No consumption fixation was observed. The recommendation retains passive opt-in collection and ends display evaluation, keeping the small display seam for reproducibility. This is not a causal finding: there is one run per task/condition, partial measurements, guidance bundled with counters, clarification-pair unblinding and a response-delay deviation.

Includes isolated fixtures, native run evidence, independently frozen outcome and fixation reviews, overhead records, and reconciled documentation. Preserves upstream behavioral guidance and collector accounting; consolidates recommendations and keeps fixtures outside the installed plugin. Repeated measured display snapshots are 137–140 characters; exact injected token cost is unknown.

Test Plan

• Full npm test passes, including 64 Mantra assertions.
• npm run validate:plugins and strict Claude marketplace/Mantra validation pass.
• Isolated Codex plugin installation succeeds at 0.9.0; fresh skill probes fired on both hosts.
• Live Claude probes verify mid-run display freshness and native usage accounting; Codex collection remains fixture-validated.
• Ada independently reviewed outcomes, fixation, integration, token efficiency and documentation; relative links and frozen source hashes verified.

Acceptance Criteria

• Verified mantra: collect trustworthy resource measurements with explicit coverage #184 measurements and distinct disabled/collect/display modes.
• Observational feedback without scores, quotas, countdowns or automatic stops.
• Outcomes, missing requirements, rework, interactions, consumption and fixation recorded separately.
• Implementation/evaluation overhead recorded, with unknown costs explicit.
• Reproducible evidence, limitations and provisional recommendation supplied.
• Negative findings retained; a cheaper incorrect result is not a success.