Defer annotations that subscript non-generic VM classes (Python 3.12/3.13 compatibility)

[proboscis]

Companion to #510 (which carries the same fix for doeff-time/time_queue.py); surfaced while locally verifying the audit-fix PRs (#508–#511) on Python 3.12/3.13.

Mechanism

doeff_vm pyclasses (DoExpr/Program, Ok) are not runtime-generic. Before Python 3.14's lazy annotations (PEP 649), Program[int] / Ok[...] raise TypeError: type '...' is not subscriptable wherever the annotation is actually evaluated:

• module-level dataclass fields (doeff_gemini/types.py → any import of doeff_gemini, which tests/test_llm_multi_provider_handlers.py does → collection Interrupted),
• module-level function signatures (packages/doeff-vm/tests/test_traceback_pure_python_frames.py → collection Interrupted),
• test-local def ... -> Program[int] signatures (tests/effects/test_pure_effect.py → 5 test-time failures).

All are masked on 3.14, so the 3.14-only local/CI habit never saw them. Fix: from __future__ import annotations in each file.

Measured (integration state = main + #491 + #508–#511)

• Before: 3.12/3.13 pytest -q → collection Interrupted (2 errors); after collection fixes: 5 failed / 1077 passed; after this PR's full set: targeted file 6 passed on 3.12 and 3.13, full-suite rerun in progress at time of writing — final numbers in the PR comment below.
• Python 3.10/3.11 remain untestable at the workspace level: uv sync refuses because the workspace's effective requires-python is >=3.12 (doeff-docker/doeff-ml-nexus) — a workspace-level decision, out of scope here.
• The Python Compatibility workflow (the only pytest CI) is manually disabled; re-enabling it is the durable guard once this and Return a program-wrapped handler from sim_time_handler and the spawned Task from ScheduleAt (#499, #503) #510 land.

🤖 Generated with Claude Code

[proboscis]

Full-suite numbers after this PR's fixes (integration state = main + #491 + #508–#511 + the #510 time_queue fix), local run:

• Python 3.12.13: pytest -q → 1082 passed, 86 skipped, 1 xpassed, 0 failed (56s)
• Python 3.13.12: pytest -q → 1082 passed, 86 skipped, 1 xpassed, 0 failed (62s)

(The 1 xpass is the known stale xfail marker that #493's fix makes pass — see #508/#510 follow-up.)

Baseline before the fixes on the same state: collection Interrupted on both versions (2 errors), and 5 runtime TypeError failures in tests/effects/test_pure_effect.py once collection was repaired.