spike: Phase 4 — measure parallel rendering, decide against sharding - #21

Note: This replaces #13, stacked on clean2/phase-3-hosting instead of clean/phase-3-hosting, and squashed on a base where the spike entrypoint's environment-variable validation already uses the require_env helper, so the squash commit itself never contains the shell idiom the scanner misread as a secret.

Replaces #5, closed. Same content, rebuilt as a single squash commit on clean/phase-4-sharding because a secret scanner flagged disposable test credentials in the old commit history and history cannot be rewritten in this environment. See the closing comment on #5 for detail.

Phase 4 of Apus was gated on a spike. It ran, it came out negative, and this PR records that plus the resulting decision. No sharding is implemented, deliberately.

The question

Can two BlueMap processes render disjoint region sets of the same map into the same map storage without damaging the zoomed-out views?

Render state is stored granularly per tile and chunk, so disjoint regions do not collide there. But lowres levels are built by aggregation — a lowres tile averages colour, height and light across a group of higher-resolution tiles, and those groups span region boundaries. Earlier research called conflicts "unlikely" because lowres tiles are also stored granularly. That reasoning conflates two different things, which is why this was measured rather than assumed.

The measurement

A single-pass reference render of the whole world, against two concurrently running containers rendering disjoint, adjacent region sets into a second bucket.

• 7 of 24 lowres tiles diverge from the reference, reproduced identically three times.
• One tile drops from 99 % rendered terrain to 91 % empty.
• A sequential control run — no race at all — corrupts 10 of 24 tiles. Order dependence confirms the mechanism independently of timing.

Granular storage prevents corruption of individual tiles. It does not prevent two shards overwriting the same aggregated lowres tile.

Full report with setup, raw numbers and an honest assessment of what the experiment does and does not cover: docs/superpowers/spikes/2026-08-09-lowres-sharding-spike.md.

The decision

The spec offered two fallbacks. This picks the second — vertical scaling via render-threads — over the two-stage alternative (shards render hires only, a final pass builds lowres):

• The two-stage route needs a custom runner bound to BlueMap-Core. Section 1.4 rules that out for the MVP, and 2.1 explains why: BlueMap-Core is not a stable public API.
• Vertical scaling already exists and costs nothing.
• No world has yet taken long enough to justify the work.

The architecture stays sharding-capable — the region list is in the bundle manifest, BlueMapMap.spec.shards exists. If a world ever does take too long, the two-stage approach is what to evaluate, and this spike is the ground to build on. Until then shards stays at 1.

Limits of the experiment, as stated in the report

Small test world, render-mask used to split rather than the scheduleMapUpdateTask API a real implementation would use, and only one split topology tested.