## Motivation

Save performance degrades as session complexity grows — on a session
with 86 panes and 78 windows, a save takes ~20 seconds.

The bottleneck is that `pane_full_command` calls the save command
strategy script once per pane. With the default `ps` strategy, each
call runs `ps -ao "ppid,args"` to dump the entire system process table,
then greps it for a single PPID. This is repeated once per pane.

Additionally, `dump_panes` piped into a while loop, running the loop
body in a subshell and preventing precomputed data from being shared
across iterations.

## Changes

**Batch the `ps` lookup into a single upfront call:**

Added `_load_ps_table` which runs `ps -ao "ppid,args"` exactly once at
the start of `save_all`, parsing the output into a bash associative
array keyed by PPID. `pane_full_command` now does an O(1) in-memory
hash lookup per pane instead of spawning a new process. Non-`ps`
strategies (pgrep, linux_procfs, gdb) are unaffected and fall through
to the original per-pane subprocess path.

**Use process substitution in `dump_panes`:**

Changed `dump_panes_raw | while ...` to `while ... done < <(dump_panes_raw)`
so the loop runs in the current shell rather than a subshell, giving it
access to the `_ps_table` associative array.

## Results

Benchmarked on a session with 86 panes across 78 windows:

| | Wall time |
|---|---|
| Before | ~20.4s (avg of 4 runs) |
| After  | ~0.70s (avg of 5 runs) |

~29× faster. The improvement scales with pane count — the more panes,
the greater the benefit.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>