Increment C is built, and the docs stop saying nothing is live

[turner]

Documentation only — no code changed.

pgb read ?format=bands off the live server on 2026-09-02 and drew from it, and the server stayed on main afterwards instead of going back to release. Neither fact was written down, and several documents asserted the opposite.

What was wrong

document	said	says
tube-map-pipeline.html	"None of this is live yet", 43 commits waiting	live since 2 Sept; the 120 s numbers are history
seqtubemap-roadmap.md	Live: Nothing; 59 commits waiting	Live: A, B and C; release is 71 behind
perf/seqtubemap-latency.md + .html	A, B and half of C undeployed, so the live numbers stand	the baseline is historical; no after column exists yet
perf/increment-b.md	confirming production numbers needs a deploy	it needs someone to read the log
perf/local-endpoint-harness.md + .html	release — the branch the server follows	past tense, isolation claim intact

Re-measured, not copied forward

perf/band-payload-sizes.mjs was re-run after #66 took the segment outlines out of the header. The roadmap's table was pre-#66 in two rows:

• 8.0 kb region: 1.10 MB against 9.97 MB of SVG, 9.0× — was 1.25 MB and 8.0×
• largest region: 9.3× — was 9.0×

New

docs/perf/increment-c.md, C's counterpart to increment-b.md: what the response became, what deliberately did not move (#23 changed no bytes; #24 changed none on the SVG route; #66 moved 4–8 numbers per golden by ≤1.5e-16), the day pgb read it, and the three corrections that came back the other way while pgb's reader was written from the spec rather than the encoder — #67, #70, #66. Each was a place the format document and the encoder had already disagreed, and each was caught by a person. That is the argument for #25.

Amended at the premise, not rewritten

Two documents were built on "the server follows release, so merging is not shipping", which is no longer true.

releasing.md gets a banner recording why the server is on main, which is not drift: one server, one port, no second port available at the hosting facility, so the only way to exercise a change against real graph data is to point the live server at main for a window. That is an argument for keeping the branch — a testing window needs a way home. It also names the immediate hazard: release points at a commit nobody has run since August, so the rollback procedure as documented would replace working code with the 120-second, request-blocking version. The procedure itself is untouched; whether to switch the server back, retire the branch for tags, or promote is a human decision.

ADR 0002 gets a dated amendment marking the two passages that no longer hold, including its own consequence about a deploy validating a batch — which came true at 71 commits with no promotion message naming them.

Checked

• All six SVGs in tube-map-pipeline.html parse.
• Every relative link in the changed Markdown resolves.
• The payload table is a fresh run, not a transcription.

The tag increment-c marks f9b05f6, the state pgb read from.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FWNvEAujjzoxiM2RCLUV7W