Run more end-to-end tests against sessions v2

[fazxes]

Summary

• FX_SESSIONS_V2=1 now reaches fx in every tmux test. Panes on the shared tmux server took that server's environment, so with the switch set the interactive suites still ran on v1. When the switch is unset, the pane has it unset too.
• With the switch set, tui-resume finds sessions under ~/.fx/sessions/v2 and expects the resume hint and the Ctrl+G relaunch to keep --sessions-v2. 47 of its 67 tests now pass on v2; the other 20 read or write v1 session files directly.
• New v2 tests for flows that had none: a resumed fx ask turn keeps the session's language; fx ask compacts on its own and resumes from the summary; a piece over 256 KB is kept as a blob and sent whole on resume; the app resumes a compacted session; text typed while the app answers steers that turn and is saved inside it; a cancelled reply is saved as cancelled; the app talks to the user while a child runs; a parent killed while a named child runs a tool records that work interrupted; a named child keeps its history after the app is killed; and a hosted terminal keeps running through an app crash, and the resumed app reaches it and stops it.
• Fault tests for the app, ACP and subagents: killed between turns, killed while a tool runs, a torn last line, a flipped byte, a missing blob, a read-only session, a second process on the same session, and a full disk. Each checks what the host says, that the log is left as it was, and that the session resumes afterwards. A damaged child log fails that message and the parent turn goes on.
• A fault grid crosses every way a run can leave a session (committed, cancelled, killed mid-reply, killed mid-tool) with every fault (none, a torn last line, a flipped byte, a missing blob, a read-only folder, a full disk, a second process) and every way back in (fx ask by id and by last, the app by id and with -c, ACP load and resume), plus a new session in each host. A refusal must name its cause and leave the log as it was, a fault that clears must give way, and a session that comes back must show the model the history the run left. On macOS the full disk is a real disk image filled to the last byte; elsewhere a file-size limit stands in.
• With FX_TRACE_SCOPES=wiring, v2 writes one trace line for each session step a host takes: the store opening and closing, a session opening and closing, a turn starting and ending, and a child starting and finishing. Each line names its session and process, so a run's session steps can be checked in order.
• The full-disk tests now size their limit from the shell's own block. Bash, which is /bin/sh on macOS, counts 1024-byte blocks and dash counts 512, so the old fx ask limit left twice the room it meant to on macOS.
• Running the suites this way found two bugs, fixed in the PRs that own the code: the picker froze for two seconds on a session open in another fx (Add the experimental v2 session store to interactive fx #1073), and resuming an interactive session with a shell result hung (Store ACP sessions in sessions v2 #1079).

Nothing changes when FX_SESSIONS_V2 is unset.

Stacked on #1082.