v1.2: Flatshare, and the demo in its original form

Flatshare v1.2. The project, called Leaseholder through v1.1, is renamed after what it does: a voice agent and coding agents sharing one GPU, the engine making room for the realtime tier. No measurement changed: every committed record keeps the names it was recorded under, and the reports and figures regenerate from them.

What changed since v1.1

• The name. The repository is StevenWang-CY/flatshare (the old URL redirects). The crates, the Python packages, the command-line tools, the systemd units, the metrics, the environment variables and the hosts' paths all carry it: flatshare-gateway, flatshare-sim, flatshare_bench, flatshare.target, FLATSHARE_*, ~/flatshare. The code still reads the records' former names, which the glossary lists, and the runbook says how a host deployed before the rename moves.
• The demo, in its original form. The README's animation and the demo video draw the dashboard of the v1.0 demo, a panel for each engine side by side, in the docs' figure style: the wait's clock and the wait against the SLO, each reply's V2V as a bar against the SLO, the turn's waveforms on its time axis with the wait between them, and the agents' tokens a second on one scale for both panels. The animation is 1,680 pixels wide in each theme, with a phone layout that sets the panels one above the other.
• The docs, consistent with the records. A pass over the README, the report and every doc for names, numbers and tense now that the A100 run is in, and for organization. Where a doc said otherwise than the records, it now says what they give: the box's 17% running share, set beside the A100's measured 6%, is the simulator's on the box's measured costs (W5's points did not sample it), as the A100 run's report now says too; the A100 agents' TTFT p50 moved by −1.2 to +1.8 s; the Colab host had 167 GiB of RAM.
• The limit of calls, tested. The behavior suite on vLLM's real scheduler now runs the model behind the docs' limit of overlapping replies: with the 48-token window, six agents keep 82% of stock's prompt tokens beside a reply every 200 steps, 9% beside one every 50 and none beside one every 40, about one reply a second at the box's step times.
• The tunnel's former keys. deploy/authorize-tunnel.sh also takes the entries of cyan's tunnel keys under their former names off a GPU node.
• The packages are versioned 1.2.0.

Assets

• flatshare-demo.mp4: the demo video, 1080p60 with sound: ten turns of the same caller on tuned vLLM 0.32 and with #63350 and #63351, and a closing card with both GPUs' runs.

The A100 run's raw records stay in v1.1, and W5's, FB's and the demo's in v1.0.