feat(config, conversation, cli): Compact only oversized items

[JeanMertz]

A compaction rule reaches every item in its turn range. That is the
right default for old turns, but it is the wrong tool for the case that
actually breaks a conversation: one fs_read_file that returned a 10 MB
log and now dominates the context window. Cutting it today means either
compacting a whole range of turns you wanted to keep, or editing
events.json by hand.

The reasoning and tool-call policies now accept an over size
threshold, so a rule reaches only the items large enough to matter:

jp conversation compact --tools=sres --over 1mb
jp query -k 'r,over=16kb+t=sres,over=1mb:..-3'

[[conversation.compaction.rules]]
tool_calls = { policy = "strip-responses", over = "1MB" }

Both surfaces compose with the existing range flags, so a threshold
narrows what a rule touches rather than replacing how it is scoped. This
also enables a tiered rule set: drop anything genuinely huge wherever it
sits, and strip everything older than the last few turns regardless of
size.

The threshold is stored in the compaction event and evaluated during
projection rather than resolved into a fixed set of matched items at
creation. Size comparison is cheap and deterministic, so it is a lazy
policy in RFD 064's terms, and projection stays a pure function of the
stored stream.

Sizes are judged per half for Strip, so a call with a short request
and a huge response loses only the response. Omit removes whole pairs,
so it is judged on the two halves combined and can never leave an
orphaned request behind. Measurement uses decoded content (tool
arguments are base64 at rest) against the raw stream, so re-running a
rule reaches the same items. summary takes no threshold: it replaces
its whole range rather than selecting from it.

Because a threshold reaches an unpredictable subset of its range, the
timeline names what it caught instead of leaving the range line to imply
everything was compacted:

Would have compacted turns 2..14 (13 total, tool responses over 1MB).
  turn 4  fs_read_file (response)  10.3 MB
  turn 9  cargo_test (response)  2.1 MB

A rule without a threshold does not itemize, and a threshold that
matched nothing says so.

Nothing changes for a rule that sets no threshold, on disk or in
behavior: a policy without one serializes as the bare string it always
was, and the built-in defaults carry no threshold.

BREAKING CHANGE: reasoning in jp conversation show -F json

The reasoning field of a compaction entry was a boolean; it is now the
serialized policy (null, "strip", or {"policy": "strip", "over": "16KB"}), matching how tool_calls already reported itself. Without
this a consumer could read tool_calls.over but had no way to see a
reasoning threshold.

Scripts testing truthiness need updating: .reasoning == true becomes
.reasoning != null.