docs(roadmap): consolidate B-series findings (B4, B6, B7, B8 + harness)

[davydog187]

Consolidate B-series perf findings in ROADMAP.md

Pulls the lessons from the B4–B8 sweep up out of the ephemeral plan files into the strategic narrative, so they survive plan cleanup and are visible to anyone evaluating future perf work.

What changes

Status section: 1,420 → 1,705 tests, 5/29 → 6/29 suite files, date.

Done section: adds the three perf-related landings since the previous update:

• PR perf(vm): fast-path the executor dispatch loop #223 (fast-path executor dispatch — numeric arith, comparisons, string concat, get_field/set_field)
• PR perf(vm): fast-path Numeric.to_signed_int64 for in-range integers #227 (B8 — numeric narrowing fast path)
• PR build(bench): add quick mode + multi-n inputs to benchmark harness #230 (bench harness: quick mode + multi-n inputs)

Direction B section: completely rewritten. The previous bullets (B1/B2/B3) were ideas from before the B4–B8 plan files were scoped; they were never used as actual plan ids. The new structure:

• Shipped: B8, bench harness.
• Tried and deferred (with findings): B6 (hypothesis obsolete after perf(vm): fast-path the executor dispatch loop #223), B7 (tuple-copy crossover at scale), B4 (case-on-tuple-elem doesn't beat list-cons head-match on the BEAM).
• What we learned: 4 takeaways — measure against today's profile, multi-n is essential for table workloads, BEAM optimizations are subtle, immutable data structures bound how fast we can be.
• Remaining lever: B5 (compile to Erlang) with current status as paused and rationale.

What does NOT change

Direction A section is left untouched. It's also stale (lists A1/A2/A3 as "in review" but they shipped long ago) but strategic narrative is human-owned per the ship-a-plan skill convention; you'll want to update Direction A separately when you decide what its current shape is.

Companion PRs

This PR rolls up the same findings that live in the per-plan deferral PRs:

• perf(vm): split table storage into array + hash parts #229 (B7 implementation, closed unmerged)
• chore(B7): defer plan; tuple cost crossover makes large tables a regression #231 (B7 deferral)
• chore(B4): defer plan; flat-tuple+PC dispatch did not beat list-cons #233 (B4 deferral)

Those PRs are the per-plan source of truth (with full benchmark tables and the implementation discoveries). ROADMAP is the cross-plan synthesis.

Diff

 ROADMAP.md | 126 +++++++++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 110 insertions(+), 16 deletions(-)

Documentation-only; nothing to verify beyond reading.