feat(agent): structured, economical compaction (soft/force ratios + fold economics)

[esengine]

Rebase of #2406 (by @SivanCola) onto current main-v2 (after #3050/#3057 landed). The branch had drifted; I applied the net diff and re-merged the compaction core by hand, since compact() had diverged on both sides.

What it does

• Soft ratio (0.5) — between the soft mark and the trigger, emit a one-shot notice instead of compacting, so the cache-stable prefix is kept intact while the user is warned.
• Force ratio (0.9) — a high-water mark that compacts even a low-value region.
• Fold economics — skip the summarizer round-trip when the foldable region is too small to pay for the extra latency/cost (bypassed by force and by manual /compact).

Rebase notes (signature reconciliation)

• main-v2 had compact(ctx, trigger, instructions); this PR had compact(ctx, force bool). Unified to compact(ctx, trigger, instructions, force) so the force flag coexists with the existing /compact <focus> text + PreCompact-hook path and the cache-diagnostics rewrite tracking from feat(agent): cache prefix diagnostics for prompt-cache churn #3057.
• Kept main-v2's token-budgeted tail (planCompaction, parameterized with a min-message count for the single-large-message fallback) rather than reverting to the PR's message-count tail, and kept main-v2's more detailed summary prompt.
• maybeCompact now layers the soft-notice and force logic on top of main-v2's anti-re-compaction compactStuck latch.
• Adapted the threshold tests to the token-budgeted tail (a single large old message folds 1→1: the summary is installed but the message count is unchanged).

Local: go build ./..., go vet, and go test ./internal/agent ./internal/config ./internal/boot ./internal/cli ./internal/control all green.

Closes #2406

[esengine]

C:/Program Files/Git/e2e

[github-actions]

🤖 Reasonix e2e benchmark

Accuracy: 3/4 (75%) · Cache hit: 75% · Tokens: 160,894 (prompt 158,586 / completion 2,308) · Compactions: 3 · Cost: ¥ 0.0463

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
compaction	❌ fail	8	102,828	1,073	63%	3	¥ 0.0419
fix-add-bug	✅ pass	3	18,342	227	98%	0	¥ 0.0011
fizzbuzz	✅ pass	3	18,821	535	99%	0	¥ 0.0017
palindrome	✅ pass	3	18,595	473	98%	0	¥ 0.0016

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (dadbd87) · triggered by @esengine

[esengine]

C:/Program Files/Git/e2e

[github-actions]

🤖 Reasonix e2e benchmark

Accuracy: 4/4 (100%) · Cache hit: 74% · Tokens: 153,899 (prompt 151,915 / completion 1,984) · Compactions: 3 · Cost: ¥ 0.0452

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
compaction	✅ pass	8	102,684	988	63%	3	¥ 0.0416
fix-add-bug	✅ pass	3	18,368	247	98%	0	¥ 0.0012
fizzbuzz	✅ pass	3	18,681	480	99%	0	¥ 0.0016
palindrome	✅ pass	2	12,182	269	99%	0	¥ 0.0009

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (dadbd87) · triggered by @esengine