De-duplicate logical session IDs before aggregating cost and tokens

[VirtualPaul]

Summary

all_session_sources() can return more than one Claude Desktop source with the same session ID. cross_session() counts every returned source, while find_session() and deletion select the first match. This can double-count tokens/cost and makes ID-based operations ambiguous.

Tested against main at fb265549480ff0d3ad8b14b28007370539b70cf3.

Reproduction

1. Have a normal Claude trace and a Claude Desktop local-agent trace whose cli_session_id is the same but whose paths differ.

2. Run:

   import meter
   from collections import Counter
   
   sources = meter.all_session_sources()
   print([item for item, count in Counter(s["id"] for s in sources).items() if count > 1])

3. Call cross_session() and compare the aggregate with the newer/preferred source alone.

4. Call find_session(the_duplicate_id, sources) and reverse sources; the selected path changes.

In a 100-source corpus, two IDs were duplicated. One duplicated pair contained the same 18 turns, 759,845 tokens, and $3.425828 twice, so the aggregate attributed $6.851656 instead of $3.425828.

Root cause

• meter.py:676-731 de-duplicates by path (known_paths), not by logical session identity.
• meter.py:4638-4651 aggregates every discovered source.
• meter.py:767-773 returns the first ID match.
• meter.py:776-808 uses that ambiguous lookup for deletion.

Expected behavior

Each logical session contributes once to global totals, and ID-based fetch/delete operations resolve deterministically to one canonical source or reject ambiguity.

Suggested fix

1. Define a canonical identity key, preferably (provider, id), with a documented fallback for missing IDs.
2. Merge duplicate metadata and choose a canonical trace deterministically (for example: exact desktop metadata match, then newest mtime, then stable path tie-breaker).
3. De-duplicate before aggregation.
4. Make find_session() reject ambiguous identifiers unless the caller provides a path-qualified/internal source key.
5. Never delete a log selected only by an ambiguous public ID.

Regression tests

• Add two different paths with the same Claude session ID and assert cross_session() counts the logical session once.
• Reverse discovery order and assert the canonical source is unchanged.
• Assert ambiguous deletion cannot move the wrong path to Trash.
• Assert genuinely distinct provider/session pairs remain distinct.