docs: refresh design-decisions.md for Phase 3c–6e state of the world (#24)

The original doc was written somewhere around Phase 3a–3b and
never updated as the engine grew. Most glaringly:

  - Decision #10 still said "No transactions yet" — Phase 4f
    landed real BEGIN/COMMIT/ROLLBACK with snapshot-based
    rollback months ago.
  - Decision #5 ("bincode for on-disk encoding") called Phase
    3c a "future refactor" — it shipped; bincode is gone from
    Cargo.toml entirely.
  - Decision #3 ("Single-file database") promised the WAL for
    "Phase 4" as if it were distant — it's live, the sidecar
    `.sqlrite-wal` is the default commit path.
  - Decision #6 predicted an LRU page cache in Phase 3d that
    didn't actually happen (still whole-file-in-RAM today).
  - Decision #4 ("Header written last on save") said "Once WAL
    lands this becomes obsolete" — partly right; both paths
    still exist for different call sites.

Also missing: every architectural decision that shipped after
Phase 3b — cell-based encoding (3c), B-Tree structure (3d),
secondary indexes as separate B-Trees (3e), POSIX flock for
concurrency (4a/4e), WAL frames + checkpointer (4b–4d),
snapshot-based rollback (4f), engine lib-plus-bin split (2.5 /
5a), Arc<Mutex> for Send+Sync (2.5), SDKs layered over a single
C FFI shim (5b–5g), WASM crate outside the workspace, lockstep
versioning (6), two-workflow release flow (6d), and the
`sqlrite-engine` crate-name workaround (crates.io had `sqlrite`
taken).

Rewrite the whole doc:

  - Group decisions by layer (Parser / Storage / Concurrency &
    durability / Query execution / Packaging / Process) so a
    reader can find what they need without scrolling through
    chronological noise.
  - Every decision now has Decision / Why / Cost as a
    consistent structure, with optional Artifact or Design
    detail blocks for the things that matter.
  - 22 decisions total, up from 11.
  - New "Still to decide" section captures known debt: case
    sensitivity, projection expressions, composite indexes,
    bounded page cache. When one of these turns into a real
    choice with an answer it can graduate up into the
    numbered list.

Every new claim cross-checked against src/ during the rewrite:
Arc<Mutex<HashMap>> confirmed in db/table.rs; bincode confirmed
absent from Cargo.toml; crate name split confirmed in root
Cargo.toml; pager's on_disk + staged + wal_cache + read-order
precedence confirmed in pager/pager.rs; WASM workspace exclusion
comment confirmed in root Cargo.toml.

Pure docs — no engine changes.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: joaoh82