fix(budget): enforce token grants and record provider cost in the ledger

[Glen-SP]

Summary

Budget enforcement was advisory: a run could overrun its token grant with nothing
throttling it, and no neuron or USD figure ever reached a run artifact. This makes
grant admission actually gate execution, adds BudgetLedger.get_exhausted_units()
as the exhaustion probe, accumulates provider cost metering in the ledger, and
escalates a ResourceBudgetError to severity error even when
require_budget_grant is False.

Informational units (cost_usd, neurons) are tracked for reporting and never
gate execution.

Fixes #81
Fixes #85

Scope

• Affected modules or public contracts: praxist/core/{budget,execution_guards, ledgers,modeling,protocol,replay,runtimes}.py,
  codex_sdk/_events.py, research_loop/stage.py. Adds one public method
  (get_exhausted_units) and one constant (INFORMATIONAL_USAGE_UNITS).
• Compatibility considerations: runs that previously overran a grant and completed
  will now be halted at the grant boundary, which is the intended correction rather
  than a regression. Existing ledgers are read unchanged; no migration.
• Task-agnostic rationale: budget admission and cost accounting are enforced by the
  shared guard layer, independent of any task's metric or protocol.

Verification

Capability validation of the budget gate, ledger and cost accounting, using the
branch's own _write_budget_grant helper so grants are constructed exactly as its
tests construct them:

Probe	Setup	Result
A	granted tokens:100, used tokens:100	exhausted_units == ['tokens'] — guard fires
B	granted tokens:100, used tokens:50, cost_usd:99	recorded, exhausted_units == [] — informational tracked, not enforced
C	granted tokens:100, cost_usd:1.0, used cost_usd:1.0	initially ['cost_usd'] — inconsistent, fixed in this branch

Probe C found that two modules disagreed about whether informational units are
enforceable: append_usage skipped them when absent from granted_budget, but
this branch also added them to ALLOWED_BUDGET_UNITS, so a grant could legally
contain cost_usd while get_exhausted_units iterated granted_budget with no
informational exclusion — turning a reporting-only unit into a hard run halt. The
exclusion is now explicit in get_exhausted_units, with a regression test.

Full checks, run on this branch merged with current main (a clean merge — no
conflicts; main's recent commits touch none of the files above):

• ruff check praxist/ tests/ — All checks passed
• ruff format --check praxist/ tests/ — 450 files already formatted
• pyrefly check — 0 errors
• pytest tests/unit --tb=short -q — 2913 passed
• run_test_coverage.py unit --fail-under 90 --fail-under-statements 95 — passed,
  93.26% total / 95.01% statements
• run_test_coverage.py integration — passed
• pytest (full) — 3353 passed, 7 skipped, 285 subtests

Checklist

• The change is focused and does not include unrelated generated files.
• Tests cover the affected behavior.
• Documentation, templates, examples, and skills were updated where needed.
• No credentials, private task data, or research-run artifacts are included.
• New dependencies and copied assets include their source and license terms.
• I have read .github/CONTRIBUTING.md and the contribution terms in LICENSE.md.