Release 0.4.0: usage reporting, routing calibration, audit nudge, handoff + dispatch/predup hardening

[dropdevrahul]

Brings main (v0.2.1) up to v0.4.0. Five commits of context + model-cost optimization work for the nexum plugin.

Highlights

0.4.0 features

• Usage reporting — store.py record-usage CLI; cost_report.py relabels the per-tier breakdown as ESTIMATED (metered session_cost stays authoritative); nx-build.md §6a logs an estimated usage row per dispatch.
• Per-repo routing calibration — route_calibration table + record/list helpers and calib-record/calib-list CLI; nx-plan.md nudges weak routes up one tier (advisory only — never down, never moves needs-strong).
• Audit nudge — scripts/audit_nudge.py throttled SessionStart one-liner wired into hooks.json.

Reliability fixes

• predup freshness guard — a read passes through when its prior tool_calls row is older than predup_max_age_seconds (default 900; 0 = old behavior), so subagent/compaction rows no longer false-deny the parent.
• dispatch hardening — max_steps_per_dispatch 6→4 plus size-aware batching (store.partition_steps_by_size + plan_preview.py), bounded by max_dispatch_context_tokens (50k).

Earlier on this branch (0.3.x)

• Pre-emptive dedup, plan cost preview, resume nudge.
• Session handoff feature; commands renamed to nx-*; context levers hardened.
• 100k auto-handoff driven off real transcript context size.
• /nexum-implement request-cost overhaul + metered cost capture.

Verification

• python3 -m pytest tests/ -q → 327 passed
• python3 tools/check_version.py → version 0.4.0 consistent across manifests
• CHANGELOG [0.4.0] entry added (Added/Changed/Fixed)

Notes

• Usage-reporting data capture begins on the next /nx-build (the orchestrator that produced this branch predates §6a).
• The predup guard mitigates but does not fully resolve cross-context false-denies; a deeper fix (tagging tool_calls rows by originating context) is tracked as follow-up.
• No v0.3.0/v0.4.0 git tag exists yet — tagging deferred pending release-mechanism decision.