From 1ea72c7f14cf76da0dc412288283b212a7369d28 Mon Sep 17 00:00:00 2001
From: Davide Cappellini <signorsasso@gmail.com>
Date: Tue, 22 Sep 2026 01:27:59 +0200
Subject: [PATCH] TM gun round 2: base was never behind; RADIAL target beats
 Linear on bmPoint

=== TASK 1: MY PREMISE WAS REFUTED ===
I instructed the job to "fix the baseline" because an earlier measurement said the
TM gun's base did not iterate flight time like `LinearGun`. MEASURED: the new gun's
base is BYTE-FOR-BYTE `LinearGun` - 18/18 runs tie exactly, p=1.000, every per-run
row byte-identical. The "non-iterating baseline" belonged to the OLD `tsetlin.nim`,
not this gun. So no fix was needed, and the earlier inference should not have been
generalised to the new gun. (It did still align the zero-correction clamp to
LinearGun's exact [0, arena] range, and reports the old BotRadius-inset base was a
wash/marginally better at 34.2%/24.7%.)

=== TASK 2: THE RADIAL TARGET - A CONTROL-VALIDATED WIN, BUT ONLY ON bmPoint ===
Instead of the lateral (GF-bucket) component - which the linear lead already
captures - the TM now predicts the RADIAL component: will the enemy be nearer or
farther than the base prediction when our bullet arrives? A 5-class radial head
sharing the same 40-bit context and TM core; the readout advances/retards the aim
distance along the base bearing.

  under bmPath (the SHIPPED metric): STRUCTURAL NO-OP
    synthetic 8/8 exact ties, p=1.0; real 33.9%/24.1% vs Linear 34.0%/24.3%
  under bmPoint: A WIN, control-validated
    TMRadial 9.4% (6013/63785) / 5.8% (42079/726652)
    Linear   7.2% / 4.7%          overall 17/1, p=0.0001
    Tsetlin  7.0% / 4.8%          overall 15/3, p=0.0075
    shuffled 7.0% / 3.6%          early 17/1 p=0.0001; overall 18/0, p<0.0001
  radial head online accuracy 48.8% vs 19.9% shuffled chance and 36.7% majority
  -> it is CONDITIONAL learning, not a constant short-range bias.
Best config: TM_RADIAL_RANGE=60, TM_RAD_MARGIN=0.25, 5 classes.

CAVEAT THAT MATTERS: a win on `bmPoint` is NOT yet evidence of a real win. `bmPath`
is the shipped SELECTION metric precisely because it beat `bmPoint` on real hit
rate (7.43% vs 4.70%). But that A/B was about which gun to PICK, not about gun
QUALITY - a gun can be better in reality while scoring worse on the selection
metric. So this needs a LIVE test, and it is the decisive one.

=== TASK 3: REVERSAL TARGET - CLEAN NEGATIVE ===
The label positive rate is only 9.7% (rev=[24772,2673]) and the head's 86.8%
accuracy is BELOW the 90.3% majority baseline: it does not learn the positive
class at all. Hit-rate effect neutral (bmPath 19.5%/18.4% vs shuffled 19.1%/17.8%,
p=0.24/0.82). Dropped.

=== OVERALL ===
Not competitive on the shipped bmPath metric (gated GF 28.3%/22.2% vs Linear
34.0%/24.3%, p=0.0075). Better than Linear on bmPoint via TMRadial (+2.2pp early,
+1.1pp overall). Per-enemy reset exists; a fresh gun per round; NO cross-battle
persistence (the user's non-negotiable).

MEASURED LIMITATION: radial mode has a high labelMiss because aiming short
resolves BEFORE the base arrival tick, biasing training toward resolvable samples.
The metric win is label-independent. A deferred-label fix is the next refinement.
INFERRED: the mechanism is surfers being NEARER than the base prediction
(range-holding); a constant-short-offset ablation would separate a learned
short-range bias from genuine per-tick conditional prediction.
---
 common_libs/guns/tm_pattern.nim               | 193 +++++++++++++++---
 common_libs/tests/sweep_tm_pattern.nim        |  95 ++++++++-
 common_libs/tests/tm_pattern_sweep_results.md | 185 +++++++++++++++++
 3 files changed, 435 insertions(+), 38 deletions(-)

diff --git a/common_libs/guns/tm_pattern.nim b/common_libs/guns/tm_pattern.nim
index d0b6ac9..64bb008 100644
--- a/common_libs/guns/tm_pattern.nim
+++ b/common_libs/guns/tm_pattern.nim
@@ -71,8 +71,30 @@ const
   TM_TRACE_SLOTS = 1024
   POS_RING = 512
   DebugTMPattern* = false
+  ## ── radial head (Task 2) ────────────────────────────────────────────────
+  ## Radial label = (enemy radius at the BASE arrival tick) - (base fire
+  ## distance), bucketed over +/-TM_RADIAL_RANGE px. Readout advances/retards
+  ## the aim distance along the base bearing.
+  TM_RADIAL_RANGE_DEF {.strdefine.} = "60.0"
+  TM_RADIAL_RANGE* = parseFloat(TM_RADIAL_RANGE_DEF)
+  TM_RAD_MARGIN_DEF {.strdefine.} = "0.25"
+  TM_RAD_MARGIN* = parseFloat(TM_RAD_MARGIN_DEF)
+  ## ── reversal head (Task 3) ──────────────────────────────────────────────
+  ## Binary: did the enemy's heading turn direction over the flight oppose the
+  ## direction it was turning at fire time?
+  TM_REV_TURN_DEG_DEF {.strdefine.} = "10.0"
+  TM_REV_TURN_DEG* = parseFloat(TM_REV_TURN_DEG_DEF)
+  TM_REV_MARGIN_DEF {.strdefine.} = "0.0"
+  TM_REV_MARGIN* = parseFloat(TM_REV_MARGIN_DEF)
+  TM_REV_GAIN_DEF {.strdefine.} = "1.0"
+  TM_REV_GAIN* = parseFloat(TM_REV_GAIN_DEF)
 
 type
+  TmTargetMode* = enum
+    tmGF        ## round-1: lateral guess-factor bucket
+    tmRadial    ## Task 2: radial displacement bucket (aim-distance correction)
+    tmReversal  ## Task 3: binary turn reversal; flips the GF correction sign
+
   TmBits* = array[TM_NLITS, uint8]
 
   TmPatternTrace = object
@@ -81,20 +103,33 @@ type
     arrivalTick: int
     baseBearing: float
     fireX, fireY: float
+    fireHeading: float
+    fireTurn: int
+    fireDist: float
     lits:      TmBits
     votes:     array[TM_CLASSES, float]
     cache:     array[TM_CLASSES, array[TM_NCLAUSES, uint8]]
     chosen:    int
+    radVotes:  array[TM_CLASSES, float]
+    radCache:  array[TM_CLASSES, array[TM_NCLAUSES, uint8]]
+    radChosen: int
+    revVotes:  array[2, float]
+    revCache:  array[2, array[TM_NCLAUSES, uint8]]
+    revChosen: int
     warm:      bool
     alive:     bool
 
   PosSample = object
     tick: int
     x, y: float
+    heading: float
     valid: bool
 
   TmPatternGun* = object
     teams:      array[TM_CLASSES, seq[int16]]
+    radTeams:   array[TM_CLASSES, seq[int16]]
+    revTeams:   array[2, seq[int16]]
+    targetMode*: TmTargetMode
     traces:     array[TM_TRACE_SLOTS, TmPatternTrace]
     # ── history ──
     posRing:    array[POS_RING, PosSample]
@@ -117,10 +152,17 @@ type
     labelMisses*: int
     chosenHist*: array[TM_CLASSES, int]
     labelHist*:  array[TM_CLASSES, int]
+    radChosenHist*: array[TM_CLASSES, int]
+    radLabelHist*:  array[TM_CLASSES, int]
+    revChosenHist*: array[2, int]
+    revLabelHist*:  array[2, int]
+    radCorrect*, radTotal*: int
+    revCorrect*, revTotal*: int
     classCorrect*: int   ## warm predictions whose class matched the eventual label
     classTotal*:   int   ## warm predictions with a resolvable label
     lastChosen*: int
     shuffleLabels*: bool  ## control: replace the computed GF label with a random class
+    forceBase*: bool      ## measurement: ignore the TM, emit the pure LinearGun base
     debugGraphics*: bool
 
 # ── TM core (Granmo Table 2/3, corrected resource allocation) ────────────────
@@ -199,6 +241,14 @@ proc bucketToGF(c: int): float {.inline.} =
   if TM_CLASSES <= 1: 0.0
   else: float(c) / float(TM_CLASSES - 1) * 2.0 - 1.0
 
+proc radToBucket(delta: float): int {.inline.} =
+  let u = clamp(delta / TM_RADIAL_RANGE, -1.0, 1.0)
+  clamp(int(round((u + 1.0) * 0.5 * float(TM_CLASSES - 1))), 0, TM_CLASSES - 1)
+
+proc bucketToRadial(c: int): float {.inline.} =
+  if TM_CLASSES <= 1: 0.0
+  else: (float(c) / float(TM_CLASSES - 1) * 2.0 - 1.0) * TM_RADIAL_RANGE
+
 proc normDeg(d: float): float {.inline.} =
   result = d
   while result > 180.0: result -= 360.0
@@ -208,6 +258,9 @@ proc normDeg(d: float): float {.inline.} =
 
 proc initTmPatternGun*(): TmPatternGun =
   for c in 0..<TM_CLASSES: result.teams[c] = tmNewTeam()
+  for c in 0..<TM_CLASSES: result.radTeams[c] = tmNewTeam()
+  for c in 0..<2: result.revTeams[c] = tmNewTeam()
+  result.targetMode = tmGF
   result.lastTick = -1
   result.prevTick = -1
   result.currentTarget = -1
@@ -221,6 +274,8 @@ proc resetLearning*(g: var TmPatternGun) =
   ## Fresh concept: wipe every clause team and the motion history. Called when
   ## the target id changes so a new opponent starts from a cold net.
   for c in 0..<TM_CLASSES: g.teams[c] = tmNewTeam()
+  for c in 0..<TM_CLASSES: g.radTeams[c] = tmNewTeam()
+  for c in 0..<2: g.revTeams[c] = tmNewTeam()
   g.totalObs = 0
   g.hasPrev = false
   for i in 0..<3:
@@ -235,7 +290,7 @@ proc tmUpdateHistory(g: var TmPatternGun, state: WorldState) =
   g.lastTick = state.tick
   let slot = ((state.tick mod POS_RING) + POS_RING) mod POS_RING
   g.posRing[slot] = PosSample(tick: state.tick, x: state.enemyX, y: state.enemyY,
-                              valid: true)
+                              heading: state.enemyHeading, valid: true)
   if g.hasPrev and state.tick > g.prevTick:
     let dx = state.enemyX - g.prevX
     let dy = state.enemyY - g.prevY
@@ -368,25 +423,25 @@ proc tmSoftGF(votes: array[TM_CLASSES, float]): float =
   for c in 0..<TM_CLASSES:
     result += (w[c] / sum) * bucketToGF(c)
 
-proc tmChooseClass(g: var TmPatternGun, votes: array[TM_CLASSES, float]): int =
-  ## Cold gun or a flat vote vector -> straight ahead (GF = 0). Otherwise the
-  ## argmax class, but ONLY when it beats the centre class by TM_CONF_MARGIN
-  ## (fraction of TM_T); otherwise stay at the centre. This is what keeps the
-  ## gun from degrading to arbitrary buckets when the TM has no real evidence.
-  let centre = (TM_CLASSES - 1) div 2
-  if g.totalObs < TM_MIN_OBS:
+proc tmChooseAt(votes: openArray[float], centre: int, margin: float,
+                nObs: int): int =
+  ## Cold gun or no class beating the centre by `margin` (fraction of TM_T)
+  ## -> the centre. Shared by the GF, radial and reversal heads.
+  if nObs < TM_MIN_OBS:
     return centre
   var best = 0
   var bestV = -Inf
-  for c in 0..<TM_CLASSES:
+  for c in 0..<votes.len:
     if votes[c] > bestV: bestV = votes[c]; best = c
   if best == centre:
     return centre
-  let margin = (votes[best] - votes[centre]) / TM_T
-  if margin < TM_CONF_MARGIN:
+  if (votes[best] - votes[centre]) / TM_T < margin:
     return centre
   result = best
 
+proc tmChooseClass(g: var TmPatternGun, votes: array[TM_CLASSES, float]): int =
+  tmChooseAt(votes, (TM_CLASSES - 1) div 2, TM_CONF_MARGIN, g.totalObs)
+
 proc predict*(g: var TmPatternGun, state: WorldState, bulletSpeed: float):
     GunPrediction =
   inc g.predictCalls
@@ -413,21 +468,52 @@ proc predict*(g: var TmPatternGun, state: WorldState, bulletSpeed: float):
   var caches: array[TM_CLASSES, array[TM_NCLAUSES, uint8]]
   for c in 0..<TM_CLASSES:
     votes[c] = tmForward(g.teams[c], lits, caches[c])
+  var radVotes: array[TM_CLASSES, float]
+  var radCaches: array[TM_CLASSES, array[TM_NCLAUSES, uint8]]
+  for c in 0..<TM_CLASSES:
+    radVotes[c] = tmForward(g.radTeams[c], lits, radCaches[c])
+  var revVotes: array[2, float]
+  var revCaches: array[2, array[TM_NCLAUSES, uint8]]
+  for c in 0..<2:
+    revVotes[c] = tmForward(g.revTeams[c], lits, revCaches[c])
 
   let chosen = g.tmChooseClass(votes)
+  let radChosen = tmChooseAt(radVotes, (TM_CLASSES - 1) div 2, TM_RAD_MARGIN,
+                             g.totalObs)
+  let revChosen = tmChooseAt(revVotes, 0, TM_REV_MARGIN, g.totalObs)
   g.lastChosen = chosen
   inc g.chosenHist[chosen]
+  inc g.radChosenHist[radChosen]
+  inc g.revChosenHist[revChosen]
 
-  var gf: float
-  if g.totalObs < TM_MIN_OBS:
-    gf = 0.0
-  elif TM_GF_MODE == "soft":
-    gf = TM_SHRINK * tmSoftGF(votes)
+  var gf = 0.0
+  var radOffset = 0.0
+  if not (g.forceBase or g.totalObs < TM_MIN_OBS):
+    case g.targetMode
+    of tmGF:
+      if TM_GF_MODE == "soft": gf = TM_SHRINK * tmSoftGF(votes)
+      else: gf = TM_SHRINK * bucketToGF(chosen)
+    of tmRadial:
+      radOffset = bucketToRadial(radChosen)
+    of tmReversal:
+      if TM_GF_MODE == "soft": gf = TM_SHRINK * tmSoftGF(votes)
+      else: gf = TM_SHRINK * bucketToGF(chosen)
+      if revChosen == 1:
+        gf = -TM_REV_GAIN * gf
+
+  # An all-zero correction must reproduce `LinearGun` BYTE-FOR-BYTE, so use its
+  # exact aim point and its exact [0, arena] clamp rather than the
+  # BotRadius-inset clamp the corrective excursions use.
+  let usedBase = (gf == 0.0 and radOffset == 0.0)
+  var px, py: float
+  if usedBase:
+    px = f.x
+    py = f.y
   else:
-    gf = TM_SHRINK * bucketToGF(chosen)
-  let aimAngle = f.bearing + gf * mea
-  let px = state.selfX + cos(aimAngle) * f.dist
-  let py = state.selfY + sin(aimAngle) * f.dist
+    let aimAngle = f.bearing + gf * mea
+    let aimDist = f.dist + radOffset
+    px = state.selfX + cos(aimAngle) * aimDist
+    py = state.selfY + sin(aimAngle) * aimDist
 
   let binIdx = tmBinForSpeed(bulletSpeed)
   if binIdx >= 0:
@@ -440,16 +526,23 @@ proc predict*(g: var TmPatternGun, state: WorldState, bulletSpeed: float):
       arrivalTick: state.tick + arrOff,
       baseBearing: f.bearing,
       fireX: state.selfX, fireY: state.selfY,
-      lits: lits, votes: votes, cache: caches,
-      chosen: chosen, warm: (g.totalObs >= TM_MIN_OBS), alive: true)
+      fireHeading: state.enemyHeading, fireTurn: g.turnSignHist[0],
+      fireDist: f.dist,
+      lits: lits, votes: votes, cache: caches, chosen: chosen,
+      radVotes: radVotes, radCache: radCaches, radChosen: radChosen,
+      revVotes: revVotes, revCache: revCaches, revChosen: revChosen,
+      warm: (g.totalObs >= TM_MIN_OBS), alive: true)
 
   when DebugTMPattern:
     echo "tmp tick=", state.tick, " bin=", binIdx, " chosen=", chosen,
-         " gf=", gf, " obs=", g.totalObs, " votes=", votes
+         " radChosen=", radChosen, " revChosen=", revChosen,
+         " gf=", gf, " roff=", radOffset, " obs=", g.totalObs, " votes=", votes
 
   GunPrediction(
-    x: clamp(px, BotRadius, state.arenaWidth - BotRadius),
-    y: clamp(py, BotRadius, state.arenaHeight - BotRadius),
+    x: if usedBase: clamp(px, 0.0, state.arenaWidth)
+       else: clamp(px, BotRadius, state.arenaWidth - BotRadius),
+    y: if usedBase: clamp(py, 0.0, state.arenaHeight)
+       else: clamp(py, BotRadius, state.arenaHeight - BotRadius),
   )
 
 proc onResult*(g: var TmPatternGun, e: FeedbackEvent) =
@@ -479,15 +572,57 @@ proc onResult*(g: var TmPatternGun, e: FeedbackEvent) =
   while delta > PI: delta -= 2.0 * PI
   while delta < -PI: delta += 2.0 * PI
   let gf = if mea > 1e-10: clamp(delta / mea, -1.0, 1.0) else: 0.0
-  let winner = if g.shuffleLabels: rand(TM_CLASSES - 1) else: gfToBucket(gf)
+
+  # The shuffled control randomises ONLY the head the current mode is claiming.
+  let shuffleGF = g.shuffleLabels and g.targetMode == tmGF
+  let shuffleRad = g.shuffleLabels and g.targetMode == tmRadial
+  let shuffleRev = g.shuffleLabels and g.targetMode == tmReversal
+
+  let winner = if shuffleGF: rand(TM_CLASSES - 1) else: gfToBucket(gf)
   inc g.labelHist[winner]
   if t.warm:
     inc g.classTotal
     if winner == t.chosen: inc g.classCorrect
 
-  for c in 0..<TM_CLASSES:
-    let d = if c == winner: 1.0 else: -1.0
-    g.teams[c].tmLearnDir(t.lits, t.cache[c], t.votes[c], d)
+  # Radial label: enemy radius at the base arrival tick minus the base fire
+  # distance. Independent of our own aim, so it is a clean target.
+  let actualRadius = hypot(g.posRing[s].x - t.fireX, g.posRing[s].y - t.fireY)
+  let radDelta = actualRadius - t.fireDist
+  let radWinner = if shuffleRad: rand(TM_CLASSES - 1) else: radToBucket(radDelta)
+  inc g.radLabelHist[radWinner]
+  if t.warm:
+    inc g.radTotal
+    if radWinner == t.radChosen: inc g.radCorrect
+
+  # Reversal label: net heading turn over the flight, opposite to the direction
+  # the enemy was turning at fire time.
+  let dh = normDeg(g.posRing[s].heading - t.fireHeading)
+  let netTurn = if dh > TM_REV_TURN_DEG: 1 elif dh < -TM_REV_TURN_DEG: -1 else: 0
+  let revWinner =
+    if shuffleRev: rand(1)
+    elif t.fireTurn != 0 and netTurn != 0 and netTurn != t.fireTurn: 1
+    else: 0
+  inc g.revLabelHist[revWinner]
+  if t.warm:
+    inc g.revTotal
+    if revWinner == t.revChosen: inc g.revCorrect
+
+  case g.targetMode
+  of tmGF:
+    for c in 0..<TM_CLASSES:
+      let d = if c == winner: 1.0 else: -1.0
+      g.teams[c].tmLearnDir(t.lits, t.cache[c], t.votes[c], d)
+  of tmRadial:
+    for c in 0..<TM_CLASSES:
+      let d = if c == radWinner: 1.0 else: -1.0
+      g.radTeams[c].tmLearnDir(t.lits, t.radCache[c], t.radVotes[c], d)
+  of tmReversal:
+    for c in 0..<TM_CLASSES:
+      let d = if c == winner: 1.0 else: -1.0
+      g.teams[c].tmLearnDir(t.lits, t.cache[c], t.votes[c], d)
+    for c in 0..<2:
+      let d = if c == revWinner: 1.0 else: -1.0
+      g.revTeams[c].tmLearnDir(t.lits, t.revCache[c], t.revVotes[c], d)
   inc g.totalObs
   inc g.trainCalls
   t.alive = false
diff --git a/common_libs/tests/sweep_tm_pattern.nim b/common_libs/tests/sweep_tm_pattern.nim
index d12d976..4cbacb6 100644
--- a/common_libs/tests/sweep_tm_pattern.nim
+++ b/common_libs/tests/sweep_tm_pattern.nim
@@ -24,6 +24,17 @@ import gun_harness/offline_range
 import range_guns
 import guns/tm_pattern
 import guns/linear
+import guns/lead_forecast
+
+type
+  LinearInsetGun = object  ## pre-fix TMPattern base: same forecast, BotRadius clamp
+
+proc predict*(g: var LinearInsetGun, state: WorldState, bulletSpeed: float): GunPrediction =
+  let f = forecastLinear(state, bulletSpeed)
+  GunPrediction(x: clamp(f.x, BotRadius, state.arenaWidth - BotRadius),
+                y: clamp(f.y, BotRadius, state.arenaHeight - BotRadius))
+
+proc onResult*(g: var LinearInsetGun, e: FeedbackEvent) = discard
 
 const repoRoot = currentSourcePath().parentDir.parentDir.parentDir
 const fixturesDir = repoRoot / "tools" / "fixtures"
@@ -33,6 +44,7 @@ type
     obs, labelMiss, traceMiss: int
     labHist, choHist: array[TM_CLASSES, int]
     classCorrect, classTotal: int
+    radCorrect, radTotal, revCorrect, revTotal: int
 
   Adapt = object
     h100, n100, h300, n300, hall, nall, f100, m100: int
@@ -47,14 +59,21 @@ type
     r: Adapt
 
   VariantKind = enum
-    vLinear, vTsetlin, vTmpat, vTmpatShuf
+    vLinear, vLinearOld, vTsetlin, vTmpat, vTmpatShuf, vTmpatBase,
+    vTmpatRad, vTmpatRadShuf, vTmpatRev, vTmpatRevShuf
 
 proc variantName(v: VariantKind): string =
   case v
   of vLinear: "Linear"
+  of vLinearOld: "LinearOldClamp"
   of vTsetlin: "Tsetlin"
   of vTmpat: "TMPattern"
   of vTmpatShuf: "TMPatternShuf"
+  of vTmpatBase: "TMPatternBase"
+  of vTmpatRad: "TMRadial"
+  of vTmpatRadShuf: "TMRadialShuf"
+  of vTmpatRev: "TMReversal"
+  of vTmpatRevShuf: "TMReversalShuf"
 
 proc loadRounds(path: string): seq[RoundSpan] =
   let dir = path.parentDir
@@ -77,6 +96,10 @@ proc addAdapt(dst: var Adapt, src: Adapt) =
   dst.st.traceMiss += src.st.traceMiss
   dst.st.classCorrect += src.st.classCorrect
   dst.st.classTotal += src.st.classTotal
+  dst.st.radCorrect += src.st.radCorrect
+  dst.st.radTotal += src.st.radTotal
+  dst.st.revCorrect += src.st.revCorrect
+  dst.st.revTotal += src.st.revTotal
   for c in 0..<TM_CLASSES:
     dst.st.labHist[c] += src.st.labHist[c]
     dst.st.choHist[c] += src.st.choHist[c]
@@ -128,6 +151,10 @@ proc replayRound(states: seq[WorldState], lastSeen: seq[int], enemyId, baseTick:
   res.st.traceMiss = after.traceMiss - obsBefore.traceMiss
   res.st.classCorrect = after.classCorrect - obsBefore.classCorrect
   res.st.classTotal = after.classTotal - obsBefore.classTotal
+  res.st.radCorrect = after.radCorrect - obsBefore.radCorrect
+  res.st.radTotal = after.radTotal - obsBefore.radTotal
+  res.st.revCorrect = after.revCorrect - obsBefore.revCorrect
+  res.st.revTotal = after.revTotal - obsBefore.revTotal
   for c in 0..<TM_CLASSES:
     res.st.labHist[c] = after.labHist[c] - obsBefore.labHist[c]
     res.st.choHist[c] = after.choHist[c] - obsBefore.choHist[c]
@@ -157,15 +184,18 @@ proc replayFixture(fx: Fixture, path: string, driver: GunDriver, metric: BulletM
 
 proc emptyStats(): GunStats = GunStats()
 
-proc makeTmpatDriver(seed: int, shuffle: bool):
+proc makeTmpatDriver(seed: int, shuffle: bool, forceBase = false,
+                     mode = tmGF):
     tuple[driver: GunDriver, gun: ref TmPatternGun] =
   let g = new(TmPatternGun)
   g[] = initTmPatternGun()
   g[].shuffleLabels = shuffle
+  g[].forceBase = forceBase
+  g[].targetMode = mode
   if seed >= 0: randomize(seed)
   result.gun = g
   result.driver = GunDriver(
-    name: (if shuffle: "TMPatternShuf" else: "TMPattern"),
+    name: (if forceBase: "TMPatternBase" elif shuffle: "TMPatternShuf" else: "TMPattern"),
     predictCb: proc(state: WorldState, bulletSpeed: float): GunPrediction =
       g[].predict(state, bulletSpeed),
     resultCb: proc(e: FeedbackEvent) = g[].onResult(e),
@@ -179,6 +209,10 @@ proc tmpatStats(g: ref TmPatternGun): GunStats =
   result.choHist = g[].chosenHist
   result.classCorrect = g[].classCorrect
   result.classTotal = g[].classTotal
+  result.radCorrect = g[].radCorrect
+  result.radTotal = g[].radTotal
+  result.revCorrect = g[].revCorrect
+  result.revTotal = g[].revTotal
 
 proc fixtureSet(name: string): seq[string] =
   case name
@@ -235,9 +269,15 @@ proc main() =
       for tok in a[11..^1].split(','):
         case tok.strip()
         of "linear": variants.add vLinear
+        of "linear_old": variants.add vLinearOld
         of "tsetlin": variants.add vTsetlin
         of "tmpat": variants.add vTmpat
         of "tmpat_shuf": variants.add vTmpatShuf
+        of "tmbase": variants.add vTmpatBase
+        of "tmrad": variants.add vTmpatRad
+        of "tmrad_shuf": variants.add vTmpatRadShuf
+        of "tmrev": variants.add vTmpatRev
+        of "tmrev_shuf": variants.add vTmpatRevShuf
         else: discard
   let metric = if metricName == "point": bmPoint else: bmPath
   let names = fixtureSet(set)
@@ -245,11 +285,13 @@ proc main() =
   echo "variant,fixture,seed,h100,n100,h300,n300,hall,nall,f100,m100,rounds,obs,labelMiss,traceMiss"
 
   var rows: seq[Row]
+  var revLabTab = initTable[string, array[2, int]]()
+  var radLabTab = initTable[string, array[TM_CLASSES, int]]()
   for name in names:
     let (fx, path) = resolve(name)
     let fxName = path.extractFilename.replace(".jsonl", "")
     for v in variants:
-      let nIter = if v == vLinear: 1 else: nSeeds
+      let nIter = if v in [vLinear, vLinearOld, vTmpatBase]: 1 else: nSeeds
       for seed in 1..nIter:
         var drv: GunDriver
         var gun: ref TmPatternGun
@@ -257,16 +299,32 @@ proc main() =
         case v
         of vLinear:
           drv = makeDriver("Linear", LinearGun())
+        of vLinearOld:
+          drv = makeDriver("LinearOldClamp", LinearInsetGun())
         of vTsetlin:
           let pair = makeTsetlinDriver(seed = seed)
           drv = pair.driver
-        of vTmpat, vTmpatShuf:
-          let pair = makeTmpatDriver(seed = seed, shuffle = (v == vTmpatShuf))
+        of vTmpat, vTmpatShuf, vTmpatBase, vTmpatRad, vTmpatRadShuf,
+           vTmpatRev, vTmpatRevShuf:
+          let mode =
+            case v
+            of vTmpatRad, vTmpatRadShuf: tmRadial
+            of vTmpatRev, vTmpatRevShuf: tmReversal
+            else: tmGF
+          let shuf = v in [vTmpatShuf, vTmpatRadShuf, vTmpatRevShuf]
+          let pair = makeTmpatDriver(seed = seed, shuffle = shuf,
+                                     forceBase = (v == vTmpatBase), mode = mode)
           drv = pair.driver
           gun = pair.gun
           obsCount = proc(): GunStats = tmpatStats(gun)
         let r = replayFixture(fx, path, drv, metric, emptyStats(), obsCount, maxRounds)
         rows.add Row(variant: variantName(v), fixture: fxName, seed: seed, r: r)
+        if gun != nil:
+          if variantName(v) notin revLabTab: revLabTab[variantName(v)] = [0, 0]
+          if variantName(v) notin radLabTab:
+            radLabTab[variantName(v)] = default(array[TM_CLASSES, int])
+          for c in 0..<2: revLabTab[variantName(v)][c] += gun[].revLabelHist[c]
+          for c in 0..<TM_CLASSES: radLabTab[variantName(v)][c] += gun[].radLabelHist[c]
         echo &"{variantName(v)},{fxName},{seed},{r.h100},{r.n100},{r.h300},{r.n300}," &
              &"{r.hall},{r.nall},{r.f100},{r.m100},{r.rounds},{r.st.obs},{r.st.labelMiss},{r.st.traceMiss}"
 
@@ -282,8 +340,8 @@ proc main() =
          &"{a.h300},{a.n300},{rateStr(a.h300, a.n300)},{a.hall},{a.nall}," &
          &"{rateStr(a.hall, a.nall)},{a.st.obs},{a.st.labelMiss},{a.st.traceMiss}"
 
-  # ── label vs chosen class histogram (TMPattern only) ──
-  for v in [vTmpat, vTmpatShuf]:
+  # ── label vs chosen class histogram (all TM variants) ──
+  for v in [vTmpat, vTmpatShuf, vTmpatRad, vTmpatRadShuf, vTmpatRev, vTmpatRevShuf]:
     if v in variants:
       let a = pooled[variantName(v)]
       var ls, cs: string
@@ -291,7 +349,14 @@ proc main() =
         ls.add &"{a.st.labHist[c]},"
         cs.add &"{a.st.choHist[c]},"
       echo &"\n# class histogram {variantName(v)}: labels=[{ls}] chosen=[{cs}] " &
-           &"onlineAcc={a.st.classCorrect}/{a.st.classTotal}"
+           &"onlineAcc={a.st.classCorrect}/{a.st.classTotal} " &
+           &"radAcc={a.st.radCorrect}/{a.st.radTotal} " &
+           &"revAcc={a.st.revCorrect}/{a.st.revTotal}"
+      if variantName(v) in revLabTab:
+        var rl, rdl: string
+        for c in 0..<2: rl.add &"{revLabTab[variantName(v)][c]},"
+        for c in 0..<TM_CLASSES: rdl.add &"{radLabTab[variantName(v)][c]},"
+        echo &"#   label hist {variantName(v)}: radial=[{rdl}] rev=[{rl}]"
 
   # ── per-run distributions (a "run" = one fixture × one seed) ──
   # Linear is deterministic: replicate its one row per fixture across seeds so a
@@ -313,6 +378,18 @@ proc main() =
         for s in 2..nSeeds:
           byVariant["Linear"][(row.fixture, s)] = byVariant["Linear"][(row.fixture, 1)]
           byVariantAll["Linear"][(row.fixture, s)] = byVariantAll["Linear"][(row.fixture, 1)]
+  if vLinearOld in variants:
+    for row in rows:
+      if row.variant == "LinearOldClamp":
+        for s in 2..nSeeds:
+          byVariant["LinearOldClamp"][(row.fixture, s)] = byVariant["LinearOldClamp"][(row.fixture, 1)]
+          byVariantAll["LinearOldClamp"][(row.fixture, s)] = byVariantAll["LinearOldClamp"][(row.fixture, 1)]
+  if vTmpatBase in variants:
+    for row in rows:
+      if row.variant == "TMPatternBase":
+        for s in 2..nSeeds:
+          byVariant["TMPatternBase"][(row.fixture, s)] = byVariant["TMPatternBase"][(row.fixture, 1)]
+          byVariantAll["TMPatternBase"][(row.fixture, s)] = byVariantAll["TMPatternBase"][(row.fixture, 1)]
 
   echo "\n# ── per-run early-rate distribution (mean / min / max, n runs) ──"
   echo "variant,earlyMean%,earlyMin%,earlyMax%,overallMean%,overallMin%,overallMax%,n"
diff --git a/common_libs/tests/tm_pattern_sweep_results.md b/common_libs/tests/tm_pattern_sweep_results.md
index a8484ea..79c41ac 100644
--- a/common_libs/tests/tm_pattern_sweep_results.md
+++ b/common_libs/tests/tm_pattern_sweep_results.md
@@ -179,3 +179,188 @@ corrective excursion net-positive. Not directly measured.
    range-correct on surfers) is where a learned correction could plausibly pay.
 4. **Richer context.** 46% accuracy leaves room; the current context lacks the
    enemy's own recent GF history / segmentation that KNN/DecayGF exploit.
+
+---
+
+# ROUND 2 — fix the base, then try a target Linear cannot predict
+
+Date: 2026-09-22. Artifacts: `common_libs/guns/tm_pattern.nim` (extended),
+`common_libs/tests/sweep_tm_pattern.nim` (extended). Raw outputs:
+`/tmp/tm2_real_path_s3.txt`, `/tmp/tm2_real_point_s3.txt`,
+`/tmp/tm_best_path_s3.txt`. All numbers below are MEASURED unless a line says
+INFERRED.
+
+Round 1's "best" gun predicted the LATERAL GF bucket. Round 2 adds a RADIAL
+head (aim-distance correction) and a binary REVERSAL head (flip the GF sign),
+both on the same 40-bit context and the same TM core, selected by a runtime
+`targetMode` (`tmGF` | `tmRadial` | `tmReversal`). The shuffled-feedback control
+now randomises only the head the active mode is claiming.
+
+## Task 1 — the base is EXACTLY Linear (premise refuted)
+
+`tm_pattern`'s base is `forecastLinear`, which already iterates the flight time
+(5-iteration fixed point, same as `LinearGun`). The only deviation from
+`LinearGun` was the wall clamp: the base path clamped to `[BotRadius, W-BotRadius]`
+(17 px inset) instead of `LinearGun`'s `[0, W]`. Added a `forceBase` flag and a
+`TMPatternBase` variant, and made the zero-correction path return `f.x, f.y`
+with the exact `[0, W]` clamp.
+
+Real DrussGT fixtures, bmPath, seeds=3, 18 fixture×seed runs, 77 rounds pooled:
+
+| variant | early | overall |
+|---|---|---|
+| Linear | 34.0% (6358/18715) | 24.3% (58297/239943) |
+| LinearOldClamp (pre-fix base, BotRadius inset) | 34.2% (6350/18592) | 24.7% (59217/239891) |
+| **TMPatternBase (forceBase, exact Linear clamp)** | **34.0% (6358/18715)** | **24.3% (58297/239943)** |
+
+Paired sign test Linear vs TMPatternBase: **18 ties, 0 wins each, p=1.000** on
+both early and overall; every per-run row (hits, shots, per-bin) is byte-for-byte
+identical. Per-run means identical: early 35.52%, overall 26.86%.
+
+Verdict: **the base was never behind.** It IS `LinearGun` to the last floating
+point. The earlier "one-shot, non-iterating baseline" finding belonged to the OLD
+`guns/tsetlin.nim`, not to `tm_pattern`. The clamp fix is a wash (the old inset
+was marginally BETTER on overall: 24.7% vs 24.3%), so there is **zero baseline
+headroom** to recover: the entire deficit vs Linear is the TM's corrective
+excursions.
+
+## Task 2 — radial target: a structural no-op under bmPath, a real WIN under bmPoint
+
+**Structural fact (from `virtual_bullets.nim`, INFERRED then confirmed):** under
+`bmPath` a bullet flies along the aim RAY until it leaves the arena; the aim
+distance only sets `fireDist` (used for the tie-break probe), it does NOT change
+the ray. Moving the aim point radially along the base bearing therefore cannot
+change a `bmPath` hit. Confirmed exactly: on the synthetic set, `TMRadial` vs
+`Linear` scored **8/8 exact ties, p=1.000** under bmPath.
+
+Under `bmPoint` the bullet resolves when `travelDist >= fireDist`, so the aim
+distance selects the arrival tick — the radial degree of freedom is live.
+
+### bmPath (shipped), real, default config, seeds=3
+
+| variant | early | overall |
+|---|---|---|
+| Linear | 34.0% | 24.3% |
+| TMRadial | 33.9% (19038/56203) | 24.1% (173577/719860) |
+| TMRadialShuf | 33.8% (18976/56085) | 24.3% (175154/719781) |
+
+Per-run means (n=18): Linear 35.52/26.86; TMRadial 35.44/26.70; Shuf 35.40/26.92.
+Paired sign tests: Linear vs TMRadial early 13/5 p=0.096; overall 15/3 p=0.0075
+(a tiny systematic LOSS, traceable to the `BotRadius` clamp perturbing the ray
+near walls when `radOffset != 0`). TMRadial vs TMRadialShuf early 8/8 p=1.000.
+
+**Verdict bmPath: no gain.** Radial is a structural no-op; the shipped metric
+therefore cannot reward Task 2.
+
+### bmPoint, real, default config, seeds=3
+
+| variant | early | overall |
+|---|---|---|
+| Linear | 7.2% (1480/20498) | 4.7% (11277/241423) |
+| Tsetlin (default gun) | 7.0% (4403/62617) | 4.8% (34588/724717) |
+| **TMRadial** | **9.4% (6013/63785)** | **5.8% (42079/726652)** |
+| TMRadialShuf (control) | 7.0% (4329/61691) | 3.6% (26116/724594) |
+
+Per-run means (n=18): Linear 13.61/5.77; Tsetlin 13.13/5.82; TMRadial
+15.32/7.06; Shuf 12.86/4.34. Paired sign tests:
+
+* **TMRadial > Linear: early 14/4 p=0.0309; overall 17/1 p=0.0001.**
+* **TMRadial > TMRadialShuf: early 17/1 p=0.0001; overall 18/0 p<0.0001.**
+* TMRadial > Tsetlin: early 17/1 p=0.0001; overall 15/3 p=0.0075.
+* TMRadialShuf vs Linear: early 10/8 p=0.81; overall 12/6 p=0.24 (control sits
+  at baseline).
+
+Online accuracy of the radial head: 48.8% (511949/1049453) vs **19.9%** shuffled
+chance under bmPoint (46.6% vs 20.0% under bmPath). Radial label histogram
+(raw, seeds=1) = [9058, 5392, 10305, 2236, 1112]: strongly asymmetric — surfers
+are often NEARER than the base constant-velocity prediction at the arrival tick
+(the base overshoots range on range-holders), so class 0 (aim 45–60 px short)
+dominates. That is the mechanism behind the win.
+
+Caveat (MEASURED): `labelMiss` is much higher for radial mode (~4.3 M vs ~1.7 M
+for GF) because aiming SHORT resolves the bullet before the base arrival tick,
+so the arrival-tick ring sample is not yet recorded. The radial head is trained
+only on resolvable samples; the win is nonetheless measured on the metric, which
+is label-independent. A deferred-label fix would be the next refinement.
+
+## Task 3 — binary reversal: not learnable, and the flip is a no-op
+
+Label: net heading turn over the flight opposes the direction the enemy was
+turning at fire time (threshold 10°). Readout: train GF as in round 1, and if
+reversal is predicted, negate the GF correction (`TM_REV_GAIN=1.0`).
+
+Label base rate (real, seeds=1, 1 round/fixture): `rev=[24772, 2673]` → the
+positive class is only **9.7%**. The head scores 86.8% (23030/26517) — **below
+the 90.3% majority-class base rate**, i.e. it is not detecting reversals at all,
+only predicting "no reversal". (The shuffled control is 50.2% because its labels
+are balanced.)
+
+### bmPath (shipped), real, default config, seeds=3
+
+| variant | early | overall |
+|---|---|---|
+| Linear | 34.0% | 24.3% |
+| Tsetlin (round-1 measurement) | 22.3% | 20.3% |
+| TMReversal | 19.5% (11062/56704) | 18.4% (132356/719725) |
+| TMReversalShuf | 19.1% (10925/57063) | 17.8% (128450/720126) |
+
+Per-run means: TMReversal 23.45/20.32; Shuf 22.51/20.01. Paired: TMReversal vs
+Shuf early 12/6 p=0.238; overall 8/10 p=0.815 → **no learning effect on hits.**
+
+### bmPath, best config (margin=0.25, shrink=0.5), real, seeds=3
+
+| variant | early | overall |
+|---|---|---|
+| Linear | 34.0% | 24.3% |
+| TMPattern (gated GF, round-1 best) | 28.3% (15815/55902) | 22.2% (159674/719742) |
+| TMReversal (gated GF + flip) | 28.4% (15944/56067) | 22.3% (160163/719761) |
+| TMReversalShuf | 28.7% (16146/56195) | 21.7% (155973/719841) |
+
+Paired: TMPattern vs TMReversal early 11/7 p=0.481, overall 7/11 p=0.481 — the
+flip changes nothing. TMReversal vs Shuf overall 13/5 p=0.096 (not significant).
+
+**Verdict: clean negative.** The reversal target as defined is too rare to learn
+(head below the majority baseline), and using it to flip the GF sign is neutral
+to slightly negative on hits. Do not pursue this label; if revisited, balance the
+positive class (per-tick reversal events, or predict the arrival turn direction
+rather than "a reversal happened").
+
+## Round-2 overall verdict
+
+* On **bmPath (the shipped metric): the TM is NOT competitive with Linear.**
+  Base = Linear exactly; radial is a structural no-op; gated GF is significantly
+  worse (28.3%/22.2% vs 34.0%/24.3%, p=0.0075); reversal does nothing. The linear
+  lead is already the best aim DIRECTION on these surfers and every learned
+  angular excursion loses.
+* On **bmPoint: the TM now BEATS Linear and the default Tsetlin gun.**
+  `TMRadial` (radial head, `TM_RADIAL_RANGE=60`, `TM_RAD_MARGIN=0.25`, 5 classes,
+  gated), 9.4%/5.8% vs Linear 7.2%/4.7% (overall 17/1, p=0.0001) and vs Tsetlin
+  7.0%/4.8% (overall 15/3, p=0.0075), with its shuffled control at 7.0%/3.6%.
+  This is the first configuration in the whole TM effort that beats both
+  baselines with a control-validated margin.
+* **Learning vs controls:** radial head 48.8% vs 19.9% chance (bmPoint); GF head
+  reproduces round 1 (48.5% vs ~20%); reversal head does not beat majority.
+
+Best configuration if the arrival-time metric is what matters: **TMRadial**. Best
+configuration under the shipped bmPath: **do nothing — keep the Linear base**. The
+evidence says the next step under bmPath is not another bucket target but either a
+richer DIRECTION representation (segmentation / pattern matching, as KNN and
+DecayGF use) or a metric that exposes the radial degree of freedom.
+
+Per-enemy specialisation / freshness (MEASURED, unchanged from round 1): each
+offline round is replayed with a FRESH gun instance and the gun calls
+`resetLearning` if the target id changes mid-battle. The offline fixtures are
+single-target, so the mid-battle reset never fires there; its effect is untested
+by these numbers. There is no persistence across battles.
+
+## MEASURED vs INFERRED (round 2)
+
+* MEASURED: every table, per-run mean, paired sign test, online accuracy, label
+  base rate, and the exact `TMPatternBase`/`Linear` byte-for-byte identity.
+* MEASURED: the bmPath radial no-op (synthetic exact ties; real bmPath tiny
+  clamp-induced loss).
+* INFERRED: that bmPath ignores radial distance because it flies a ray — read
+  from `virtual_bullets.nim`, then confirmed by the synthetic tie.
+* INFERRED: that the radial win comes from surfers being NEARER than the base
+  prediction (range-holding), supported by the asymmetric radial label histogram
+  but not separately modelled.