feat(compaction): report user turns the retention budget could not hold

[esengine]

Follow-up to #8419, which made compaction keep folded user turns verbatim within
a budget. This closes the two gaps that PR left open.

The blind spot

Retention shipped with no way to see it fail. A turn past the budget still folds
into the digest, and the projection reads exactly the same whether every
constraint survived or one was dropped. So the single case where compaction
still hands a user's own words to the summarizer was the single case nobody
could observe — the same shape of failure the original change set out to fix,
one level up.

It also left [[keep]] useless in practice. An escape hatch documented only in
SPEC.md does not help the person who needed it at the moment they needed it.

Change

keepUserTurns returns a userTurnRetention count that rides the existing
telemetry path into CompactionTelemetry as user_kept / user_dropped, and a
committed checkpoint that had to drop a turn emits a warning naming
[[keep]]. The notice fires after commit and never on a rejected candidate — a
checkpoint that was not installed folded nothing, so warning there would be a
false alarm.

Two structural cleanups ride along, both in service of the change rather than
beside it:

• The retention code moves to compact_user_turns.go. compact.go was at 705
  lines against the 800-line ceiling, and this is one responsibility.
• partitionFoldForProjection drops its early / carried return values,
  which have been nil since content-driven summary replaced the early-user-turn
  hoist. Threading retention through would otherwise have made it five results.

Coverage

Effect test at the final boundary (internal/boot). Drives the real Build
assembly through an actual fold and asserts the constraint reaches the provider
request — with a summarizer whose digest deliberately records nothing, so a
digest cannot be what carries it. The constraint is the second user turn on
purpose: the first is pinned into the stable prefix and would not exercise
fold-region retention at all.

This test was verified by mutation, not just by passing: stubbing
keepUserTurns to return early turns it red with constraint absent from the post-compaction request. Restored and re-verified green.

Sub-agent inheritance. Sub-agents build from subagentOptions, the single
construction point task / read_only_task / parallel_tasks children share,
which already forwards ContextWindow, KeepPolicy and CompactRatio — so
retention is inherited with no new code. A test pins that inheritance and the
child-scoped budget, because a parent's instructions are the child's
constraints and nothing else in the child transcript records them.

Conservation law. partitionCoversRegion now also asserts that every user
turn in a region is either kept or counted as dropped. A turn can no longer go
missing from the accounting the way it previously could from the projection, and
every test using that helper inherits the check.

Verification

gofmt -l .        clean
golangci-lint     0 issues
repolint          113571 — byte-identical to the main-v2 baseline
go test ./...     all green, zero failures

Cache-impact: low - the stable system-prompt prefix is untouched; this adds counters to existing compaction telemetry and one post-commit notice. No change to which messages compose the projection, so the provider-visible bytes for a given transcript are identical to #8419.
Cache-guard: internal/boot/compaction_retention_effect_test.go asserts the post-compaction provider request through the real Build assembly (mutation-verified), plus the unchanged internal/agent/cachehit_e2e_test.go.
System-prompt-review: esengine - the only internal/boot change is a new test file; no system-prompt text, tool schema, memory content, or boot wiring is modified.
Documentation-impact: updated - docs/SPEC.md "What survives a fold" now states that a drop is reported rather than silent, with the matching text in docs/SPEC.zh-CN.md.