fix(store): keep fork resets isolated from parent tape

[Andy963]

Summary

Fix ForkTapeStore.reset() so resets performed inside a fork stay isolated from the parent tape until merge-back.

Previously, reset() always called both the in-fork store reset and parent.reset(). That meant tape.reset inside a fork could immediately wipe the parent tape even when merge_back=False, breaking fork isolation and risking data loss.

What Changed

• Track the active fork tape and whether it was reset within the current fork context.
• In ForkTapeStore.reset():
  • keep immediate parent reset behavior outside a fork
  • keep immediate parent reset behavior for non-fork tapes
  • defer parent reset for the active fork tape
• In fetch_all():
  • hide parent entries once the active fork tape has been reset in the current fork
• In fork() finalization:
  • apply parent.reset(tape) only when merge_back=True and the forked tape was reset
  • then append forked entries back to the parent

Why

This restores the expected isolation semantics:

• merge_back=False: fork changes, including reset, must not affect the parent tape
• merge_back=True: reset should replace parent tape contents only when the fork is committed

Tests

Added regression coverage for:

• reset inside fork with merge_back=False preserves parent entries
• reset inside fork with merge_back=True replaces parent entries
• reset inside fork hides parent entries during reads
• reset outside fork still resets the parent immediately

Verification

• uv run ruff check .
• uv run mypy src
• uv run pytest -q

[frostming]

That meant tape.reset inside a fork could immediately wipe the parent tape even when merge_back=False, breaking fork isolation and risking data loss.

Can we fix this by not wiping the parent tape when merge_back=False?

[Andy963]

That meant tape.reset inside a fork could immediately wipe the parent tape even when merge_back=False, breaking fork isolation and risking data loss.

Can we fix this by not wiping the parent tape when merge_back=False?

Yes, and this PR implements that in a more complete way.

Instead of only skipping the parent wipe when merge_back=False, it keeps reset fork-local during execution and only applies the parent reset when the fork is merged back. That preserves isolation semantics and avoids partial behavior differences between read-time and merge-time state.