fix(usage): cost display honesty — sub-cent labels, cost buckets, included notes

[kshitijk4poor]

Summary

Three cost-display honesty fixes. Small, cheap, kills a real confusion class.

1. Sub-cent cost label rendering (#79220)

Problem: cost labels formatted at 2 decimal places, so for models priced under ~$1/Mtok (DeepSeek, etc.) a normal turn costing $0.004640 rendered ~$0.00 despite amount_usd carrying full Decimal precision.

Fix: format_cost_label(amount) scales precision to magnitude:

• Zero → $0.00
• Sub-cent (< $0.01) → ~$0.0046 (4 dp)
• Below $0.00005 → ~$<0.0001 (never a zero-looking 4dp truncation)
• Normal → ~$1.23 (2 dp, unchanged)

Shared by per-response labels (estimate_usage_cost) AND the insights aggregate formatters, so sub-cent honesty can't regress on one surface while fixed on the other.

2. Cost bucket surfacing in insights (#77223)

Problem: aggregate cost views summed estimated_cost_usd and reported a single number. Sessions with cost_status = 'included' (subscription-included providers like openai-codex) contribute zero — correct per-session, but invisible in the overview. In the reporter's DB, 315 of 473 sessions were included and didn't show at all.

Fix: both formatters now display three buckets:

  💰 Cost
  ────────────────────────────────────────────────────────
  Estimated:         ~$3.75
  Included:           315 session(s) (subscription — no provider invoice)
  Unknown:            3 session(s) (no pricing data)

Gateway: **Cost:** ~$3.75 estimated | 315 included (subscription) | 3 unknown

3. Subscription-included cost notes

estimate_usage_cost attaches a "subscription-included; no provider invoice for usage" note to CostResult for subscription-included routes, so consumers can distinguish "free because subscription" from "free because $0 pricing".

Validation

Check	Result
tests/agent/test_usage_pricing.py + test_insights.py	62 passed / 0 failed
E2E probe: sub-cent label via real estimate_usage_cost route	PASS (~$0.0046 observed)
E2E probe: all 7 estimate_usage_cost callers tolerate the note	PASS (none parse .label)
E2E probe: repo-wide grep for 2dp-label consumers	PASS (no hits — nothing breaks on 4dp)
Mutation check: reverting the formatter routing makes the sub-cent aggregate guard fail	PASS
ruff	clean

Fixes #79220
Fixes #77223

[alt-glitch]

This was generated by AI during triage.

Duplicate of #77247 for the included/estimated/unknown Insights cost-bucket implementation. This PR also bundles the separate sub-cent label work from #79220; please consolidate or split the overlapping bucket implementation before review.

[kshitijk4poor]

Good flag — looked at #77247 closely. They overlap in the formatter hunks but aren't duplicates:

• The included_cost_sessions / unknown_cost_sessions counters already exist on main (_compute_overview, insights.py:494-508) — neither PR introduces them.
• This PR is display-layer only: it surfaces the existing counters in the terminal/gateway formatters and fixes the sub-cent $0.00 label class (Cost label formats at 2dp, so sub-cent per-turn costs render as $0.00 #79220), sharing one format_cost_label across both surfaces.
• feat(insights): surface included/estimated/unknown cost buckets (#77223) #77247 goes further on the computation side: it adds _estimate_included_market_cost — a new at-market USD valuation of subscription-included sessions — which this PR deliberately does not attempt.

So the resolution is: this PR lands the display surfacing + label honesty; #77247's market-value estimation remains valuable as an enhancement on top (its formatter hunks will need a small rebase over this, but the computation half is untouched). Not consolidating them into one PR keeps the display fix small and the valuation feature separately reviewable.