2026-08-19 - Crunch continued: phases 2, 0e, and 3

146 tests, gate green. D054 and D055 added.

Phase 2 (loading half). Symbol databases load from disk with their own suffix, and SymbolDb::explains measures a name list against real hashes - the self-verifying loop from D025, where a collision is the proof. orbistoun-cli verify reports the fraction of a module's imports a database can name. The generator half is still outstanding and needs a suffix to test against.

Phase 0e finished. orbistoun-cli report surveys a module, persists a run report, and prints the delta from the previous run of the same title. Retention purges on startup. Verified end to end: first run has no diff, second run diffs against it, and a different title correctly reads as a first run because identity is the content hash.

Phase 3 done and verified on both platforms. orbistoun-cli load reserves the span a module demands: a 96 KiB module and a 96 MB commercial executable both place cleanly at a supplied base.

Surprises, all from running rather than reasoning.

• Per-segment reservation was wrong (D054). Windows reserves at 64 KiB granularity, so segments a few pages apart collided with each other - three of five "failing" on a module that fits fine. A module is one contiguous span. Real segment vaddrs are also not page-aligned (0x147f0), so the span rounds outwards.
• VirtualAlloc2 placeholders are unnecessary. Plain VirtualAlloc at an explicit base already refuses rather than overwriting, which is the whole requirement.
• The executable links at vaddr 0, so it needs a placement base exactly as a module does. The attempt failed with "requested 0x0, kernel returned 0x134a2340000" - and that refusal is the design working, not a bug.
• The Linux path was completely broken (D055) and only running it in the multipass VM showed that: MAP_PRIVATE was missing, so every reservation returned EINVAL. Worse, every mmap error was mapped to Conflict, so it read as "range taken" and would have sent a reader hunting a phantom occupant. Wrong error messages are worse than none.
• The rustdoc gate caught stale docs - orbistoun-mem still described itself as unimplemented after being implemented. Cheap catch, and exactly why that lint is on.

Next. Phase 0b (ABI spike) to de-risk phase 4, then phase 4 itself.