Weekly recap: include skill auto-pull bytes in Y (v1.1 follow-up)

The savings recap in `src/notifications/sources/local-usage.ts` (shipped in `feat/hivemind-savings-recap`) currently computes:

```
Y = memorySearchBytes / 4
Z = 0.7 · Y (per the LoCoMo 1.7× benchmark)
```

`memorySearchBytes` covers two of hivemind's four context-injection channels into a session:
1. ✅ Memory grep results (`Bash(grep ~/.deeplake/memory/summaries/...)` tool_result bytes)
2. ✅ Sessions raw grep results (`Bash(grep ~/.deeplake/memory/sessions/...)` tool_result bytes — same path filter)
3. ❌ **Skill auto-pull content** (this issue) — bytes of `~/.claude/skills//SKILL.md` files that skillify pulled, loaded into context by Claude Code's built-in skill loader at session start.
4. ❌ Excluded by design: SessionStart instructions block (hivemind's "DEEPLAKE MEMORY: ..." preamble — it's hivemind explaining itself, not user past content).

Currently the headline is a conservative lower bound — real savings (including skill contribution) is higher than the displayed number.

## To include skill bytes in Y

One of:

- **(a) Transcript scan**: walk the session JSONL for `attachment.type === "hook_success"` entries whose stdout contains skill-loader system-reminders; count the bytes of the SKILL.md content the loader injected. Most precise (counts only skills Claude actually loaded this session), but requires understanding Claude Code's skill-loader system-reminder format and could break across CC versions.

- **(b) Filesystem scan**: at SessionEnd, `ls ~/.claude/skills/*/SKILL.md` filtered by mtime in the session window AND `author: hivemind` in YAML frontmatter (or equivalent provenance marker), sum their byte sizes. Simpler, but counts skills that were *pulled-but-not-used* this session. Probably suffices for v1.1 — still under-counting compared to "true value delivered," which is fine for the marketing claim.

Recommendation: ship (b) first as a follow-up PR. Add a `skillBytes` field to `UsageRecord`, sum it into Y at render time. Same 1.7× multiplier applies.

## What changes downstream

- `src/notifications/usage-tracker.ts`: add `skillBytes: number` to `UsageRecord`. Backward-compat read for older records.
- `src/hooks/session-end.ts`: in `recordSessionUsage`, scan `~/.claude/skills/` after parsing transcript, populate `skillBytes` on the record.
- `src/notifications/sources/local-usage.ts`: `Y = (memorySearchBytes + skillBytes) / 4`.
- Tests: extend `transcript-parser.test.ts` and `usage-tracker.test.ts` to cover the new field.

## Verification

After this lands, the recap headline should jump (real savings > current under-counted display). Verify by:
1. Add a new skill via `hivemind skillify pull` so a SKILL.md lands under `~/.claude/skills/`.
2. Open a session — note the recap headline value.
3. Open a NEW session — recap should refire with a HIGHER number than before (skill bytes now contribute to Y).

## Related

- PR shipping v1: feat/hivemind-savings-recap on activeloopai/hivemind