wallet costs: align estimate contexts with upload's real provider/context selection

[nijoe1]

wallet costs currently approximates the next upload: it builds --copies storage contexts by reusing the wallet's first N active datasets and padding with new-dataset placeholder objects (cast to StorageContext; prepare()'s costing reads only dataSetId/withCDN in the pinned SDK).

The real upload path instead runs selectHealthyProviders() (reachable, unique providers) and createContexts() (matching provider + source + CDN metadata), so the estimate can diverge:

• it may reuse two datasets on the same provider, though a multi-copy upload requires unique providers;
• it may reuse a dataset whose CDN/source metadata upload would not reuse;
• it prices arbitrary historical datasets rather than the providers the next upload will select.

Divergence changes dataset-creation fees, CDN lockups, effective monthly rate, and depositNeeded. The plain-object placeholders also depend on SDK costing internals that a compatible minor upgrade could change.

Interim state (shipped in PR #30): the CLI/README/skill describe the command as an approximate requested-copies estimate — not the source of truth — and note that upload re-quotes via its own prepare() before spending.

Acceptance:

• The estimate builds contexts the same way upload does (health-checked unique provider selection + metadata-matched dataset resolution), or uses SDK support for costing hypothetical contexts without side effects.
• No plain-object as StorageContext casts.
• Regression: estimate for a wallet with same-provider duplicate datasets and mismatched-CDN datasets matches what upload's own prepare() quotes.
• Restore the stronger wording in README/skill once aligned.

Context: PR #30 follow-up review (2026-07-23), "wallet costs is still not the exact upload quote it claims to be".