feat: establish unified Macintosh runtime foundations

[benletchford]

Summary

• document the refined one-process architecture, ownership laws, audited gap matrix, migration order, and deletion gates
• add a single address-routing authority for both CPU memory contracts, including atomic mixed and read-only scalar and bulk writes
• add CPU-free task-owned continuation identity and lifecycle validation, plus transactional cross-ISA completion and callback timing fixes
• converge TickCount, ordinary NewHandle, and Trap Manager walking slices on shared process authorities
• require explicit system-memory provenance for privileged permanent trap chains on both ABI paths
• add a CI ratchet that prevents adapter state, trampoline machinery, and direct native-call construction from growing

Audit status

This is the foundation increment for the tracking issue, not the end of the migration. The migration contract records the remaining concrete work: kernel-owned context banks, resumable MenuSelect and MDEF effects, one CopyBits transfer engine, manager dependency waves, and the final MacintoshProcess composition root. The tracking issue remains open until those deletion gates are complete.

Verification

• cargo check --lib --no-default-features
• cargo check --all-features
• cargo test --lib --no-default-features --features test-support -- --quiet: 4,932 passed, 3 ignored, 0 failed after rebasing onto public master
• scripts/check-unified-runtime-guardrails.sh
• shell syntax and diff whitespace checks

Part of #1363.

[rlanday]

Heads-up: while re-baselining perf work against current master we found two host-work regressions that trace to this PR (git log -S), and we will send fixes.

1. TrapDispatcher::trap_table_address now calls bus.protected_code_ownership(), which clones readonly_code_ranges into a ProtectedCodeOwnership on every trap dispatch. On a 3 in Three headless replay (200 M guest instructions, sample at 1 ms), trap_table_address is 22% of self time and the memmove under it another 6%, plus the allocator traffic (madvise, bzero). Before this PR the lookup was two bus reads. The fix is a borrowed check (bus.protected_code_contains(address)) for the immutable-bus callers; the snapshot stays for install_trap_address, which needs the mutable bus.
2. MacMemoryBus::route is now on every read_word/read_long and shows as 6.5% of self time out of line; we are looking at a fast path for the flat-RAM case.

Scale, with #1220 rebased on master so its single-step effect is excluded: master executes about 3× the host instructions of the pre-#1364 tree on 3 in Three (312 B vs 103 B for the same 200 M guest instructions, identical ticks and screenshots) and 3.8× on SimCity 2000 (192 B vs 51 B per 400 M). A bisect over 3c63ca4..c40588f is running to confirm the split between this PR and #1362; the profile attribution above is from the symbol level.

[rlanday]

Follow-up with the fixes and the confirmed attribution.

Bisect (3 in Three headless replay, 200 M guest instructions, host instructions retired, single runs): 5d87791 = 105.9 B; eef89a3 (this PR) = 268.0 B; 305027d (#1362) = 308.6 B; 83af901 = 308.5 B.

Fixes, all against master with #1220 rebased on it so the single-step mode is excluded:

• perf(trap): look up trap table addresses without copying or scanning the protected-code ranges #1419: trap_table_address no longer clones the protected-code ranges per dispatch, and the range list is kept sorted and merged so is_protected_head, readonly_code_contains and readonly_code_overlaps are binary searches instead of linear scans over ~1,000 entries.
• perf(bus): decide flat guest accesses inline when no foreign address space is attached #1420: MacMemoryBus::route answers the flat-RAM case inline when no foreign address space is attached.

Together with #1421 (the #1362 side), machine idle, 5 pairs each, identical ticks and framebuffer hashes: 3 in Three −60.7% host instructions and −53.0% CPU time (35.4 s → 16.6 s); SimCity 2000 −39.1% and −30.8% (18.6 s → 12.9 s). The pre-#1364 tree is 105.9 B on 3 in Three; the fixed stack is 122.7 B.