refactor: unify agent history and harden runtime diagnostics

Preserve user guidance through history compression, stabilize Flash index targets, and retire withdrawn checks without applying stale verdicts.

Improve runtime diagnostics, tool responses, and history rendering. Simplify documentation and comments.

Validation: 2076 tests passed in the full suite; 381 affected tests and 26 frontend tests passed after cleanup. Ruff and the quality ratchet passed.

config/artemis.jsonc

@@ -192,7 +192,7 @@
     // ==========================================================================
     // 🌐 4.3 Shared Capabilities & Global Overrides
     // ==========================================================================
-    // 🧠 Unified Step-Memory Runtime (history-module redesign v3, M0–M5)
+    // Shared step memory and history compression
     // Transcript ledger + layered compression for the Pro operator.
     "memory": {
       // Background summarization runtime (visual/capsule lenses)
@@ -204,15 +204,33 @@
       "transcript": {
         // ROLLBACK SWITCH: set false to restore the legacy per-turn 2-message
         // prompt path byte-for-byte (L2/L3 chunk compression rides this same
-        // flag). Default on since M5 (2026-09-01) per the on-device A/B:
-        // turn latency −77% P50, cache-hit calls 43%→88%, output tokens −74%.
+        // flag). See services/token_meter.py for measured session token usage
+        // and cache-hit ratios.
         "enabled": true,
         "context_budget_tokens": 80000,
+        // Below start_ratio of the operator's own measured prompt size, closed
+        // segments only prepare their summary; the original turns are kept in
+        // context and replaced once occupancy reaches this ratio. Measurements
+        // come from the operator; other agents' calls do not affect this gate.
+        "start_ratio": 0.35,
         "soft_ratio": 0.7,
         "hard_ratio": 0.9,
         "min_active_steps": 5,
+        // Screenshot edge: 6 while the operator's measured context is below
+        // start_ratio (nothing is being replaced yet), 3 from there up. A
+        // screenshot reaching the edge is replaced in place by its visual
+        // summary (the 'Current Screenshot' label leaves with it) and never
+        // restored; the depth only decides how long new ones stay.
         "image_scrub_depth": 3,
+        "image_scrub_depth_relaxed": 6,
         "pending_grace_steps": 3,
+        // Text edge: once a newer observation exists (depth 1) the UI element
+        // list, the plan recitation and the per-turn ephemeral blocks
+        // (reminders, hints, guidance wrappers) leave the older observation.
+        // User instruction bodies remain in history. The text edge stays
+        // shallow on purpose: only the live observation carries an indexed
+        // element list, so a stale [n] index can never be picked up as a
+        // target.
         "xml_scrub_depth": 1,
         // Local historical-state hint (no model call); distance calibrated
         // 2026-09-01 from 460 dHash-stamped on-device steps (same screen <=4,
@@ -223,6 +241,10 @@
       // Segment (L2) chunk compression: three-band chunks, era folding
       "chunking": {
         "max_steps": 12,
+        // Size/pressure closes wait until at least this many turns are
+        // eligible, so heavy turns do not become one-step chunks (one capsule
+        // call each); milestone closes are exempt, the hard line waives it.
+        "min_steps": 3,
         "target_source_tokens": 2000,
         "model": "gemini-3.8-flash",
         "max_chunks": 8