Fix stale slot_1 availability bit in LineListNode

[zariuq]

Hope this isn't a waste of your time. I hardly understand it but Codex thinks it ran into and patched a small bug while working on something MORK-related. GPT-5.4 Pro and Claude Code agree it seems solid. Figure if the AIs agree, I should at least ping ya'll 🙏.

Codex's message:

This fixes a stale header invariant in LineListNode.

Bit 12 is dual-use: when slot_1 is unused, it means slot_1 is unavailable because slot_0 consumed all key bytes. Two paths could shorten slot_0 below KEY_BYTES_CNT while leaving that bit set:

• shorten_key_len::<0>
• the single-slot drop_head_dyn path

That left slot_1 empty but still marked unavailable, which could send later insertions down the fallback split path incorrectly.

This patch:

• clears the stale bit in shorten_key_len::<0>
• routes the single-slot drop_head_dyn shortening path through the same helper
• adds a debug invariant in validate_node
• adds two focused regression tests covering both local writers of the stale state

Tests:

• cargo test -p pathmap releases_slot_1_availability -- --nocapture

I also confirmed this fixes a real downstream crash in a Metamath/MORK consumer.

If you want an even shorter version:

Fix a stale slot_1 availability bit in LineListNode.

When slot_0 was shortened below KEY_BYTES_CNT, two paths could leave bit 12 set even though slot_1 was empty and reusable. That stale state could send later insertions down the fallback split path
incorrectly.

This patch fixes both writers, adds a debug invariant, and adds two focused regression tests.

Tested with:

• cargo test -p pathmap releases_slot_1_availability -- --nocapture

Also confirmed against a downstream Metamath/MORK reproducer.