fix(extract): persist CLI token usage in cost.json (#1769)

[TPAteeq]

Fixes #1769.

Problem

graphify extract calculated and printed token usage after each run, but never
persisted it to graphify-out/cost.json. The only ledger writer lived in the
agent-driven skill pipeline, so direct CLI refreshes left the original entry
and cumulative totals unchanged.

The labeling usage added in #1694 was similarly available to cluster-only and
label, but was not persisted.

Fix

• Add a shared record_cost_run() ledger writer using the existing
  cost.json schema.
• Append a run after every successful clustered or --no-cluster extraction,
  including unchanged zero-token incremental refreshes.
• Persist cluster-only and label token usage after the command completes
  successfully.
• Recompute cumulative totals from run history, repairing missing or stale
  total fields.
• Serialize concurrent writers with a cross-process lock and use atomic file
  replacement so interrupted writes cannot truncate the ledger.
• Preserve existing file permissions and respect the process umask when
  creating a new ledger.
• Treat ledger failures as telemetry warnings rather than failing an otherwise
  successful graph build.

Testing

Regression coverage includes:

• appending to an existing ledger across incremental extraction runs;
• clustered, --no-cluster, and unchanged incremental exits;
• labeling usage appended to an existing ledger;
• cumulative-total recovery from run history;
• concurrent writers;
• atomic replacement failure preserving the original file;
• existing permission preservation and new-file umask handling.

Full suite: 3093 passed, 30 skipped.

Additional checks:

• ruff check clean
• pyright graphify/cost.py clean
• python -m tools.skillgen --check clean
• graphify update . completed