docs(roadmap): correct the P2 row, two GW-2 steps, and the discovery sprint cut

[JArmandoAnaya]

Docs only, no behaviour change. Executing P2 surfaced four places where the committed docs no longer matched the code or a decision already taken.

1. The roadmap's P2 row contradicted the agreed style contract

It described road styles as "draggable assets preserving prior attributes" — the reverse of what was decided, and the same wording that made GW-2 step 11 self-contradictory. A style defines the lane profile; it cannot both apply and preserve the lane edits it replaces. The row now states the real contract: replaces the cross section and boundary marks, preserves everything orthogonal to it (reference-line geometry, elevation, junction connectivity).

Same row: (supersedes #95/#97) sat on enclosed-area surfaces, but #95 is tee-on-commit and #97 is grade-match — both extend-from-endpoint concerns. Moved.

2. The roadmap never mentioned the Help workstream

#265 and #266 are two of P2's eleven sprints and were scoped as permanent cross-pillar infrastructure every later pillar grows — yet docs/roadmap/README.md had zero references to it. A reader could not tell it existed. The P2 row now names it, and the sprint conventions record the cross-pillar workstream rule (titled for the workstream, labelled for both) so help-sM isn't read as a violation of the pN-sM convention.

3. Two GW-2 steps amended (approved)

Both are documentation defects, not implementation-dependent wording, so they land now rather than at close-out — otherwise nine sprints run against an acceptance script describing a product nobody built. #264 keeps ownership of any further refinement once the tools exist.

• Step 3 claimed "a junction is created automatically at the overlap". Nothing does that. create_road_tool.cpp:186-189 only calls create_road/create_linked_road, and SnapKind is {Grid, RoadEndpoint, TangentContinuation} — no side or crossing candidate. Junctions are made solely with the Create Junction tool. The kernel ops do exist (assembly::cross_onto_road, tee_onto_road, attach_t_junction), so P2 wires them up in p2-s3: extend-from-endpoint + tee/cross on commit #214 and the step now says when the junction forms.
• Step 11 asked for "lane count edits" to survive a style that defines the lane count. Now explicit and testable, and it names the elevation profile edited back in step 7 — which is the interesting preservation case in this workflow.

A short amendments note under ## Steps records both, following GW-1's precedent of an inline prose note rather than a separate section.

4. The discovery report's sprint cut was stale

It still listed interior add_lane under p2-s1 after that moved to p2-s2 (renumbering must remap JunctionConnection::lane_links — the contract p2-s2 rewrites, so one sprint should own it), and predated #262–#266 existing, so those rows had no issue links. Also links the dead-dependency finding to #268 (libigl never linked; Manifold unused outside a smoke test — filed, deliberately unscheduled, outside P2).

Verification

Relative links and anchors checked locally against what the docs link check job enforces (lychee --offline --include-fragments). docs/roadmap/golden_workflows/gw1_camera.md is deliberately untouched — it carries the maintainer's in-progress GW-1 hand-run.