SirRoboGarage/docs/env_reference.md

# Environment variable reference

Every knob the bot reads. **Most are read once per process, at module init; a few
are read lazily on first use** (`TR_TMHORIZON_SHIFT`/`_BIG_MULT`/`_LOG`/
`_RESET_ON_TARGET` in `ensureConfig`, `TR_PATTERN_RAD_*` inside `predict`). Either
way, changing one mid-run has no effect.

> **The bot is spawned by the server (or by the GUI if it starts its own server).**
> It inherits the **server's** environment. Export in the shell that launches the
> server/GUI — not in an unrelated terminal. See
> [Did my env vars actually reach the bot?](#did-my-env-vars-actually-reach-the-bot).

**Read these two traps before the tables — they are the two ways this document has
misled people:**

1. **Defaults are silent for numeric/bool knobs.** Unset / empty / unparseable
   means the shipped default. Only the enumerated *string* knobs
   (`TR_RACK_*`, `GUN_VBULLET_METRIC`, `GUN_SELECTOR_MODE`, `GUN_SELECTOR_RANK`,
   `GUN_SELECTOR_TIEBREAK`) **warn on stderr** on a bad value; the
   `envInt`/`envFloat`/`envBool` helpers fall back to the default with **no
   warning at all**.
2. **Some flags are presence-based, not value-based.** They are read with
   `existsEnv`, so **`TR_POWER_LOG=0` turns the log ON** (any value does).
   Presence-based: `TR_POWER_LOG`, `TR_RAM_LOG`, `TR_MOVEMENT_LOG`,
   `TR_RECORD_WORLDSTATE`, `TR_RADAR_FORCE_SPIN`, `TR_RADAR_SCANLOG`,
   `TR_TRACKER_PROBE`. Value-based (`0`/`false`/`off` really disable):
   `TR_ENV_REPORT`, `TR_TMHORIZON_LOG`, `TR_TMHORIZON_ACCURVE`,
   `TR_TMHORIZON_RESET_ON_TARGET`, `TR_POWER_POLICY`, `TR_POWER_FINISH_KILL`,
   `TR_RAM_OPPORTUNITY`, `TR_RAM_PLAN`, `GUN_SELECTOR_POOL`,
   `TR_VBULLET_ADMIT_ONLY`.

---

## COMPILE-TIME vs RUNTIME: the two namespaces

**A `-d:` compile-time define and a runtime environment variable are different
namespaces. The naming is only *partly* parallel, so guessing an env name from a
define name is unreliable.**

| compile-time define | runtime env | notes |
|---|---|---|
| `-d:TMH_NSTATES=N` | `TR_TMHORIZON_NSTATES=N` | **The ONLY pair with both forms.** Env wins at runtime; clamped `2..4096`. `tm_horizon.nim:102,139,381` |
| `-d:TMH_NCLAUSES=N` | *(none)* | compile-time ONLY — no env form. `tm_horizon.nim:101` |
| `-d:TMH_S_DEF="x"` | *(none)* | compile-time ONLY. `tm_horizon.nim:103` |
| `-d:TMH_MIN_OBS=N`, `-d:TMH_STALE_MAX=N` | *(none)* | compile-time ONLY. `tm_horizon.nim:106,108` |
| `-d:TM_NCLAUSES=N`, `-d:TM_NSTATES=N`, `-d:TM_S_DEF="x"`, `-d:TM_MIN_OBS=N`, `-d:TM_CLASSES=N` | *(none)* | `tm_pattern.nim` — compile-time ONLY. Note **`TM_NCLAUSES`** (no underscore) here. `tm_pattern.nim:50,53,55,57,59` |
| `-d:TM_N_CLAUSES=N`, `-d:TM_N_STATES=N`, `-d:TM_S_DEF="x"`, `-d:TM_T_DEF`, `-d:TM_WINDOW_SIZE=N` | *(none)* | `tsetlin.nim` — compile-time ONLY. Note **`TM_N_CLAUSES`** (underscore) here, and `TM_N_STATES` not `TM_NSTATES`. `tsetlin.nim:29,116,118,119,120` |

Rules and warnings:

- The `-d:TMH_X` ↔ `TR_TMHORIZON_X` mapping holds **only for `NSTATES`**.
  `TMH_NCLAUSES`, `TMH_S_DEF`, `TMH_MIN_OBS`, `TMH_STALE_MAX` have **no env
  form** — there is no `TR_TMHORIZON_NCLAUSES` etc.
- `-d:TM_S_DEF` is defined in **both** `tm_pattern.nim` (default `"3.0"`) and
  `tsetlin.nim` (default `"1.5"`); the same flag sets both.
- **If you set an env var and the bot seems to ignore it, the name is probably
  wrong — `TMH_NSTATES` is NOT an env var, the env var is
  `TR_TMHORIZON_NSTATES`.** Likewise `TM_N_CLAUSES`, `TMH_NCLAUSES`, etc. are
  compile-time only.

---

## Did my env vars actually reach the bot?

### 1. The spawn trap in one sentence

The bot is spawned by the **server/GUI**, so it inherits the **server's**
environment — exporting a variable in a separate terminal does **nothing**,
because that terminal is not the bot's parent.

### 2. The no-code check (works today, no boot report needed)

`/proc/PID/environ` is the environment **at exec time** — authoritative, exactly
what the bot started with, before it read anything:

```sh
for p in $(pgrep -f ModularBot); do
  echo "== pid $p"
  tr '\0' '\n' < /proc/$p/environ | grep -E '^(TR_|GUN_)' | sort \
    || echo '  (none - wrong process!)'
done
```

**Warning:** filtering to `TR_|GUN_` **hides a wrongly-named variable** — e.g. a
variable called `TMH_NSTATES` will not show up under this filter (and neither will
a `TR_TMHORIZON_*` typo). When you are debugging a *missing* var, grep for the
distinctive part of the name with **no prefix filter**:

```sh
tr '\0' '\n' < /proc/$p/environ | grep -i tmh
```

### 3. The boot report (`TR_ENV_REPORT`, default `1`)

> **Available in the build that adds the boot report** (`ModularBot_garage/src/
> env_report.nim`, committed as `9bf3005`). If your binary predates it, use §2.

The bot prints a one-shot `[env]` block to **stdout** at boot
(`/tmp/modularbot_stdout.log` when launched by the GUI, or the GUI/server
console). Recover just the report with:

```sh
grep '^\[env\]' /tmp/modularbot_stdout.log
```

It has two sections plus build identity:

- **A. raw process environment** (`[env] --- A. raw process environment ...`):
  every `TR_*`/`GUN_*` this process *actually* received, sorted, followed by
  `[env] raw: N TR_*/GUN_* of M total environment variables`. If `N == 0` it
  prints a loud
  `[env] WARNING: this process has NO TR_*/GUN_* variables - the env you exported did NOT reach the bot.`
  It then prints the process identity: `pid`/`ppid`, `cwd`, the bot's own command
  line (`self_cmd`), and the **parent's command line** (`parent_cmd`). The parent
  command line is the conclusive proof of who spawned the bot — under the GUI it
  is the server/GUI, never your interactive shell.
- **B. effective values** (`[env] --- B. effective values (resolved at boot) ---`):
  one line per knob, `[env] NAME = value  (source: env|default)`, i.e. the value
  the bot will *actually* use after parsing, clamping and fallback (e.g.
  `TR_TMHORIZON_NSTATES` is clamped to `2..4096`; the rack's empty-set fallback is
  shown). Where a value is only resolvable later it says
  `(raw - not resolved here)` rather than guessing. It also prints
  `[env] rack active 1v1 = ...` and `[env] rack active melee = ...`.
- **build identity**: `NimVersion`, compile date/time, and the binary path + size
  + mtime, so a stale binary is obvious.

`TR_ENV_REPORT=0` suppresses the report (it is value-based: `0`/`false`/`no`/`off`
all suppress).

### 4. Launching the GUI with knobs

Export in the **same shell** that launches the GUI/server, then start it from that
shell. Because the launcher **appends** to the log, clear it **first** so you read
only the new run:

```sh
rm -f /tmp/modularbot_stdout.log
export TR_POWER_ENERGY_MIN=1.0 TR_RACK_PATTERN=off
./start-gui.sh                 # whatever launches the server/GUI
grep '^\[env\]' /tmp/modularbot_stdout.log
```

### 5. Prove the trap deliberately

Terminal A — where you *exported*, but never launched anything:

```sh
export TR_POWER_ENERGY_MIN=1.0
```

Terminal B — where you launch the GUI:

```sh
./start-gui.sh
grep '^\[env\]' /tmp/modularbot_stdout.log | grep WARNING
# [env] raw: 0 TR_*/GUN_* of ... total environment variables
# [env] WARNING: this process has NO TR_*/GUN_* variables - the env you exported did NOT reach the bot.
```

Terminal A's variable never reached the bot because the bot's parent is the
server, not terminal A. Fix it by exporting in terminal B before launching.

---

## Measured verdicts — read before tuning (MEASURED on the live arena vs real DrussGT)

These are live battles, not offline replay. Do not re-derive an offline-only win
and ship it. (Offline claims are labelled as such.)

| experiment | result | verdict |
|---|---|---|
| shipped rack (Pattern only) | **24/49 rounds = 49.0%** (up from 12.5% earlier in the session) | the baseline |
| `TR_TMHORIZON_WINDOW=150` (with TMHorizon admitted) | **13/49 = 26.5%** vs 49.0%, **p=0.036** | **MEASURED HARMFUL live.** Looked like an offline +9.3pp late-accuracy win; the arena disagreed. Keep `0`. |
| TMHorizon `TR_TMHORIZON_NSTATES=2` | 42.9% | no better than shipped (p>=0.8) |
| TMHorizon `TR_TMHORIZON_NSTATES=8` | 42.9% | no better than shipped (p>=0.8) |
| TMHorizon `TR_TMHORIZON_NSTATES=64` (default) | 53.1% | not significant vs Pattern (p>=0.8) |
| refusing to fire below power 1.0 | 22/49 vs 21/49, **p=1.0** | no help |
| sub-1.0 vs >=1.0 bullet accuracy | 10.80% vs 10.35%, **p=0.37** | sub-1.0 bullets are **not** less accurate |
| `TR_POWER_POLICY=0` (uncapped) | real hit rate 10.61% → 7.88%, p=0.0012 | policy helps; keep it on |
| `TR_VBULLET_ADMIT_ONLY=1` | 87 → 146 ticks/s (+68%) | keep it on |
| `TR_MOVEMENT=tfil_ring` | round wins 16/49 → 6/49, p=0.012 | glass cannon; do not ship |
| `TR_RAM_OPPORTUNITY=1` | 0/59 opportunity→contact conversions | finisher-only is the default |

On the TM rows: **no TM configuration beats the shipped rack**, and **low inertia
does not help** (contrary to an earlier offline trend, where it was a substitute
for forgetting, not a fix). On the power rows: the floor experiment exposed the
real tradeoff — power `p` costs a `10+2p` tick reload, so a floor costs ~13% of
your shots/round, while the bullets it suppresses were *not* less accurate (they
are simply worth less damage per hit, `4p`, which is what makes low-power
shooting *look* like missing).

---

## The ones you'll actually use live

| variable | default | what it does |
|---|---|---|
| `TR_MOVEMENT` | `tfil` | movement engine: `tfil` (shipped) or `tfil_ring` (range-weighted variant; any other value falls back to `tfil`) |
| `TR_RACK_<GUN>` | see below | move a gun in/out of the rack per mode |
| `TR_POWER_POLICY` | `1` | `0` = uncapped control arm (today's behaviour without the energy policy) |
| `TR_POWER_LOG` | off | **presence-based**: if the var exists at all (even `=0`) log each power decision |
| `TR_RAM_LOG` | off | **presence-based**: log ram on/off with the reason |
| `TR_MOVEMENT_LOG` | off | **presence-based**: log movement band/class changes |
| `TR_TMHORIZON_LOG` | off | value-based: `1` = let the horizon TM gun log its thinking per shot |
| `TR_ENV_REPORT` | `1` | print the boot-time `[env]` report to stdout; `0` suppresses it |

---

## 1. Gun rack — which gun(s) may be chosen

| variable | default | what it does |
|---|---|---|
| `TR_RACK_<GUN>` | `PATTERN=both`, all other 15 guns `off` | membership: `both` \| `1v1` \| `melee` \| `off` |
| `GUN_RACK_DISABLE` | empty | comma-separated gun **ids** to remove entirely (older mechanism; the gun never even spawns virtual bullets) |
| `GUN_STATS_PATH` | `/tmp/gun_stats.jsonl` | where the per-round per-gun stats are written |
| `GUN_SHOTLOG_PATH` | `/tmp/shot_log.jsonl` | where the per-shot JSONL is written |

`<GUN>` names (ids 0-15, `selector.nim:52`):
`HEADON LINEAR TSETLIN CIRCULAR GUESSFACTOR PATTERN WALLBOUNCE ACCEL STOPSHOT
DISPLACE AVGLEAD DECAYGF KNN TMSELECT TMPATTERN TMHORIZON`.

Shipped default membership (`selector.nim:DefaultRackMembership`):

| gun | default | gun | default |
|---|---|---|---|
| HEADON | off | STOPSHOT | off |
| LINEAR | off | DISPLACE | off |
| TSETLIN | off | AVGLEAD | off |
| CIRCULAR | off | DECAYGF | off |
| GUESSFACTOR | off | KNN | off |
| **PATTERN** | **both** | TMSELECT | off |
| WALLBOUNCE | off | TMPATTERN | off |
| ACCEL | off | TMHORIZON | off |

`TR_RACK_<GUN>` accepts aliases: `both`/`any`/empty → both;
`1v1`/`single`/`lock` → 1v1; `melee`/`multi` → melee;
`off`/`none`/`disabled` → off. An **unknown value warns on stderr and falls back
to `both`** (`selector.nim:parseRackMembership`).

The `1v1`/`melee` split is driven by **server truth**, `getEnemyCount()`:
`== 1` → 1v1 rack, `> 1` → melee rack (`virtual_bullets.rackMode`).

**Empty-set fallback:** if every gun is filtered out for a mode, the *selector*
falls back to the **full rack** so it can never end up with no gun
(`admittedGuns`, `virtual_bullets.nim`). Note this fallback is on *selection*;
the spawn gate (`TR_VBULLET_ADMIT_ONLY`) uses `rackAdmitted` and would spawn
nothing, so an all-`off` rack is not a useful configuration.

**The shipped default is Pattern only.** To restore the full old rack:

```sh
TR_RACK_PATTERN=both TR_RACK_HEADON=both TR_RACK_LINEAR=both TR_RACK_TSETLIN=both \
TR_RACK_CIRCULAR=both TR_RACK_GUESSFACTOR=both TR_RACK_WALLBOUNCE=both \
TR_RACK_ACCEL=both TR_RACK_STOPSHOT=both TR_RACK_DISPLACE=both TR_RACK_AVGLEAD=both \
TR_RACK_DECAYGF=both TR_RACK_KNN=both TR_RACK_TMSELECT=both ./out/ModularBot
```

To force one gun alone: `TR_RACK_PATTERN=off TR_RACK_TMHORIZON=both`.
Enabling a gun is `TR_RACK_<GUN>=both`; TMHorizon additionally needs
`TR_RACK_TMHORIZON=both` before any `TR_TMHORIZON_*` knob can matter.

## 2. The selector (virtual-bullet fitness)

| variable | default | what it does |
|---|---|---|
| `GUN_VBULLET_METRIC` | `path` | how a virtual bullet is scored: `path` (swept ray — the better coarse signal, 7.4% vs 4.7% real) or `point` (arrival accuracy) |
| `GUN_SELECTOR_MODE` | `relative` | tie-band model: `relative` (scale-aware) or `absolute` (legacy fixed bars) |
| `GUN_SELECTOR_WINDOW` | `100` | rolling window (ticks) for the fitness rates; **clamped `1..100`** |
| `GUN_SELECTOR_TIE` | `0.20` | tie band width: tied if `rate >= bestRate*(1-this)` |
| `GUN_SELECTOR_FLOOR` | `0.25` | relative floor: fire HeadOn only if best rate < this × its peak |
| `GUN_SELECTOR_MINOBS` | `50` | min observations before a gun×bin may compete (clamped to ≥1) |
| `GUN_SELECTOR_POOL` | `on` | pool the gun's power bins when computing its rate |
| `GUN_SELECTOR_RANK` | `mean` | window statistic: `mean` \| `wilson` \| `ucb` \| `thompson` \| `shrunk` |
| `GUN_SELECTOR_SHRINK` | `20.0` | empirical-Bayes shrink strength (`shrunk` rank only; clamped ≥0) |
| `GUN_SELECTOR_DWELL` | `10` | ticks the incumbent gun is held before it may be displaced (clamped ≥0) |
| `GUN_SELECTOR_MARGIN` | `0.05` | a challenger must beat the incumbent by this fraction to displace it (clamped ≥0) |
| `GUN_SELECTOR_SEED` | unset | integer seed for the tie-break RNG; unset = time+pid per process |
| `GUN_SELECTOR_TIEBREAK` | `off` | `off` (shipped) / `point` / `pointCommit`. Measured **negative** on real hit rate; opt-in only |
| `GUN_SELECTOR_POINT_TIE` | `0.5` | margin used by the `point` tie-break (clamped `0..1`) |

Constants live in `virtual_bullets.nim:16,28,31,33,35`.

**Note:** the per-tick random draw inside the tie band is **load-bearing** —
replacing it with commitment to the virtual-best measurably *lowered* real hit
rate (7.02% → 5.10%, p=0.002). Three attempts to "smarten" the band all failed.

## 3. Power policy (energy economy)

| variable | default | what it does |
|---|---|---|
| `TR_POWER_POLICY` | `1` | `0` = uncapped control arm. **MEASURED: turning it off drops real hit rate 10.61%→7.88%, p=0.0012** |
| `TR_POWER_FAR_DIST` | `200` px | beyond this = "bad chances zone" |
| `TR_POWER_FAR_CAP` | `1.0` | power cap beyond that distance |
| `TR_POWER_MID_CAP` | `2.0` | cap when close + healthy but chances are not above average |
| `TR_POWER_REF` | `0.0` | reference probability: `0` = the gun's own mean, `>0` = a fixed threshold |
| `TR_POWER_ENERGY_HI` | `80` | at/above this energy, no energy-slope cap |
| `TR_POWER_ENERGY_LO` | `20` | at/below this energy, cap = `ENERGY_MIN` |
| `TR_POWER_ENERGY_MIN` | `0.5` | the **cap** at/below `ENERGY_LO` — **NOT a minimum-power floor** (see below) |
| `TR_POWER_ENERGY_MAX` | `3.0` | the cap at/above `ENERGY_HI` (3.0 = effectively uncapped) |
| `TR_POWER_FINISH_KILL` | `1` | cap power at the **smallest bullet that can still kill** the enemy |
| `TR_POWER_LOG` | off | presence-based; logs each decision: `[power] p=… cap=… reason=… dist=… selfE=… enemyE=… gun=…` |

Constants live in `virtual_bullets.nim:842-851`; the rule core is
`applyPowerPolicy`/`energySlopeCap`/`powerToKill`.

> **`TR_POWER_ENERGY_MIN` is a CAP, not a floor.** It is the ceiling applied when
> *our* energy is at/below `TR_POWER_ENERGY_LO`. Setting it to `1.0` raises the
> low-energy ceiling to 1.0; it does **not** forbid shots below 1.0. Sub-1.0
> bullets come from the caps (`ENERGY_MIN`, `FAR_CAP`, `FINISH_KILL`). MEASURED:
> refusing to fire below power 1.0 does not help (22/49 vs 21/49, p=1.0), and
> sub-1.0 bullets are *not* less accurate (10.80% vs 10.35%, p=0.37).

WHY the slope: bullet speed is `20-3p`, so **lower power = faster bullet** (less
lead error, higher hit chance), fires more often (`10+2p` ticks) and drains energy
more slowly (`p`/shot). `E[dE] = p(3P-1)`, so break-even is **1/3 regardless of
power** — and our measured hit rates are 5–27%, far below it. A power floor at
1.0 costs ~13% of your shots/round (the longer reload) and buys no accuracy.

WHY the finishing cap: server 1.3.1 caps the damage **score** at the energy
actually removed, so **overkill scores nothing**. Damage is `4p` (p≤1) / `6p-2`
(p>1), so the smallest killing power is `E/4` for E≤4 and `(E+2)/6` for 4<E≤16.

## 4. Movement

| variable | default | what it does |
|---|---|---|
| `TR_MOVEMENT` | `tfil` | `tfil` (shipped) or `tfil_ring` (the range-weighted variant; any other value falls back to `tfil`, no warning) |
| `TR_MOVEMENT_LOG` | off | presence-based; `1` = log band/range-class changes for the ring mover |
| `TR_TFIL_RANGE_LO` / `_HI` | `100` / `200` px | the target-range band the ring mover prefers |
| `TR_TFIL_RANGE_TEMP` | `0.4` | peakiness of the range weighting. `0` = plain uniform draw = exact control |
| `TR_TFIL_RANGE_K` | `60` | softness of the falloff outside the band |
| `TR_TFIL_CORRIDOR_HEAT` | `10.0` | heat added along a bullet's corridor to the wall |
| `TR_TFIL_WALL_HOTNESS` | `15.0` | peak wall radiance |

Defaults read in `the_floor_is_lava_ring.nim:116-133`.
**Discrepancy to be aware of:** that file's header comment still says
`CORRIDOR_HEAT default 5.0` / `WALL_HOTNESS default 10.0` (the pre-retune values);
the **code defaults are `10.0` / `15.0`** (commit `7f6ccfb`). The code is the
truth.

**Ring mover is NOT the default and should not be shipped as-is.** It had the best
**LIVE** hit rate of anything measured (20.28%, server-side event sidecar ground truth), but it
halves engagement range and **halves survival** (round wins 16/49 → 6/49, p=0.012) — a glass
cannon.

**CORRECTION (see `docs/offline_harness_trust.md`):** this line previously claimed the ring mover
had "the best offline hit rate". That was **wrong — the ring mover was never measured offline at
all**; the offline harness scores *guns*, and has no movement driver. So this was **not** an
offline-vs-live calibration failure (which is how it was repeatedly described). It was a **metric
mismatch**: a movement arm judged on hit rate instead of damage/run and round wins — and hit rate
is exactly the metric that hid its collapse.

## 5. Ramming

| variable | default | what it does |
|---|---|---|
| `TR_RAM_OPPORTUNITY` | off | enable the proactive ram. **MEASURED: 0/59 opportunity→contact conversions** — a straight-line pursuit cannot catch an equal-speed enemy |
| `TR_RAM_OPP_DIST` | `200` px | opportunity trigger distance |
| `TR_RAM_OPP_MARGIN` | `15` | how much MORE energy we must have than the enemy |
| `TR_RAM_ABORT_DMG` | `2.0` | abort an in-progress ram above this incoming **energy per turn** (the "bullet rain" abort) |
| `TR_RAM_PLAN` | off | the speculative "change of plan" trigger |
| `TR_RAM_PLAN_DIST` / `_MARGIN` / `_HITRATE` | `250` / `20` / `0.05` | its thresholds |
| `TR_RAM_LOG` | off | presence-based; `1` = log `[ram] ON/OFF` with reason |

Defaults in `ram_decision.nim:79-97`. The **finisher** ram (enemy <20 energy, we
are healthier) is always on and is the only path that converts.

## 6. The horizon TM gun (`TMHORIZON`)

**Only relevant when the rack admits it** (`TR_RACK_TMHORIZON=both`); the shipped
rack is Pattern-only, so these knobs are inert by default.

| variable | default | what it does |
|---|---|---|
| `TR_TMHORIZON_LOG` | off | value-based; `1` = per-shot thinking log + per-round summary |
| `TR_TMHORIZON_SHIFT` | `2.0` deg | how far it may move the aim off Pattern's answer. **`0` = predict but never move the aim** (isolates prediction from application) |
| `TR_TMHORIZON_BIG_MULT` | `1.5` | multiplier when the predicted correction is BIG |
| `TR_TMHORIZON_NSTATES` | `64` | **automata inertia** — the "mood". Lower = adapts faster, higher = more stubborn. Clamped `2..4096`; overrides the `-d:TMH_NSTATES` compile-time value |
| `TR_TMHORIZON_RESET_ON_TARGET` | on | wipe the model when the target changes (no-op in 1v1) |
| `TR_TMHORIZON_WINDOW` | `0` | sliding window: retrain on only the most recent N samples. `0` = keep everything. **ON is MEASURED HARMFUL live: 26.5% vs 49.0% round wins, p=0.036 — keep it at `0`.** (Offline it was +9.3pp late accuracy.) |
| `TR_TMHORIZON_RESET_DROP` | `0` | change detection: if rolling accuracy falls this many points below its own peak, treat it as "the enemy changed" and re-learn |
| `TR_TMHORIZON_ACCURVE` | off | value-based; `1` = log the rolling accuracy periodically |
| `TR_TMHORIZON_RETRAIN_EVERY` | `50` | samples between windowed retrains (clamped ≥1) |
| `TR_TMHORIZON_EPOCHS` | `1` | epochs per windowed retrain (clamped ≥1) |

Defaults in `tm_horizon.nim:126-152,381-412`.

Learning **persists across rounds** of the same battle (so it can overfit the
current enemy) and wipes only on a new battle or a target change. **No learning is
written to disk — nothing carries between battles.**

**Measured:** no TM configuration beats the shipped rack — N=2 42.9%, N=8 42.9%,
N=64 53.1% (all p>=0.8 vs Pattern), N+window 26.5%. Low-inertia `NSTATES=2` does
**not** help (contrary to an earlier offline trend).

## 7. Other gun knobs

| variable | default | what it does |
|---|---|---|
| `TR_PATTERN_RAD_OFFSET` | `0.0` px | shift Pattern's aim along its own bearing (negative = aim short) |
| `TR_PATTERN_RAD_SCALE` | `1.0` | multiplier on Pattern's aim distance |

(`pattern_matcher.nim:26-27`.) Both are **structurally incapable of changing the
shot** — the live aim is bearing-only, so a purely radial offset is invisible.
Measured byte-identical on `bmPath`. Kept for experiments; leave at defaults.

## 8. Measurement / instrumentation (all off by default, all zero-cost when off)

| variable | default | what it does |
|---|---|---|
| `TR_VBULLET_ADMIT_ONLY` | `1` | only admitted guns spawn virtual bullets. **MEASURED: +68% tick rate** (87→146 ticks/s); `0` restores the old behaviour |
| `TR_ENV_REPORT` | `1` | print the boot-time `[env]` report to stdout; `0` suppresses it (value-based) |
| `TR_RECORD_WORLDSTATE` | off | presence-based; append every tick's WorldState to `/tmp/worldstate_record.jsonl` for offline replay |
| `TR_RADAR_FORCE_SPIN` | off | presence-based; force the old stateless full-spin melee radar (for A/B on one binary) |
| `TR_RADAR_SCANLOG` | off | presence-based; append per-round radar scan/coverage JSON |
| `TR_RADAR_SCAN_LOG_PATH` | `/tmp/radar_scan_log.jsonl` | where that goes |
| `TR_TRACKER_PROBE` | off | presence-based; per-tick enemy tracker vs server enemy count |
| `TR_TRACKER_PROBE_PATH` | `/tmp/tracker_probe.jsonl` | where that goes |

The **adaptive-melee radar** has no env knobs. Its tuning lives in compile-time
constants in `radars/adaptive_melee_radar.nim:36-50`: `MaxRadarTurnRate=45`,
`FreshnessTicks=16`, `FreshStreakTicks=3`, `MarginDeg=20`,
`EnterTrackWidthDeg=270`, `ExitTrackWidthDeg=300`. Only
`TR_RADAR_FORCE_SPIN`/`TR_RADAR_SCANLOG` are runtime.

## 9. Test harness only (not read by the bot itself)

| variable | default | what it does |
|---|---|---|
| `TR_SERVER_JAR` | `/home/davide/Downloads/robocode-tankroyale/robocode-tankroyale-server.jar` (1.3.1) | which server the tests launch. Set to the 0.35.5 jar (`server_manager.LegacyServerJar`) to reproduce old numbers |
| `TR_BATTLE_RUNNER` | `/home/davide/Projects/tank-royale/runner/examples/lib/robocode-tankroyale-runner.jar` | which runner the tests launch (`runner_process.nim`) |
| `TR_BATTLE_RUNNER_DIR` | `tools/battle_runner/` | class directory form of the runner |

(`TR_SAMPLE_BOTS` is read by `SNNBot_garage/tests/test_bullet_economy.nim` only — not
by the ModularBot test harness.)

## 10. Compile-time knobs (`-d:` flags, need a rebuild)

| flag | default | file / effect |
|---|---|---|
| `-d:TMH_NSTATES=N` | 64 | `guns/tm_horizon.nim:102` — automata inertia (the runtime `TR_TMHORIZON_NSTATES` overrides it) |
| `-d:TMH_NCLAUSES=N` | 40 | `guns/tm_horizon.nim:101` |
| `-d:TMH_S_DEF="x"` | `"3.0"` | `guns/tm_horizon.nim:103` specificity (`s`) |
| `-d:TMH_MIN_OBS=N` | 24 | `guns/tm_horizon.nim:106` — cold gate: below this many samples the TM emits no correction |
| `-d:TMH_STALE_MAX=N` | 8 | `guns/tm_horizon.nim:108` — a sample is dropped if the enemy was not seen this recently |
| `-d:TM_CLASSES=N` | 5 | `guns/tm_pattern.nim:50` — GF buckets |
| `-d:TM_NCLAUSES=N` | 40 | `guns/tm_pattern.nim:53` — clauses per class |
| `-d:TM_NSTATES=N` | 64 | `guns/tm_pattern.nim:55` |
| `-d:TM_S_DEF="x"` | `"3.0"` | `guns/tm_pattern.nim:57` specificity (`s`). **Higher = LONGER clauses**, measured. `s=1.0` is degenerate |
| `-d:TM_MIN_OBS=N` | 24 | `guns/tm_pattern.nim:59` |
| `-d:TM_CONF_MARGIN_DEF` / `TM_SHRINK_DEF` / `TM_GF_MODE` / `TM_SOFT_BETA_DEF` / `TM_RADIAL_RANGE_DEF` / `TM_RAD_MARGIN_DEF` / `TM_REV_TURN_DEG_DEF` / `TM_REV_MARGIN_DEF` / `TM_REV_GAIN_DEF` | see `tm_pattern.nim:62-96` | strdefines, compile-time only |
| `-d:TM_WINDOW_SIZE=N` | 10 | `guns/tsetlin.nim:29` |
| `-d:TM_N_CLAUSES=N` | 50 | `guns/tsetlin.nim:116` — clauses per output |
| `-d:TM_N_STATES=N` | 32 | `guns/tsetlin.nim:118` — automaton range `[-N..N]` |
| `-d:TM_S_DEF="x"` | `"1.5"` | `guns/tsetlin.nim:119` — **same flag name as `tm_pattern`; one `-d` sets both** |
| `-d:TM_T_DEF` | `""` (→ `float(TM_HALF)`) | `guns/tsetlin.nim:120` |

## Names that look real but do nothing

| name | reality |
|---|---|
| `TMH_NSTATES` (as an env var) | not an env var. The runtime var is `TR_TMHORIZON_NSTATES` |
| `TR_VBULLET_METRIC` | does not exist. The metric env var is `GUN_VBULLET_METRIC` |
| `TR_POWER_LOW_ENERGY` | removed (replaced by the energy slope). Only referenced in a comment; `getEnv` never reads it |
| `TR_TRACKER_RECONCILE` | printed/set by `common_libs/tests/measure_corpse_melee.nim` only; **no bot module reads it** |

## Adding / finding knobs

```sh
# everything the code reads
grep -rn 'TR_[A-Z_]*"' --include=*.nim ModularBot_garage/src common_libs | sort -u
# what a specific knob defaults to
grep -rn 'TR_POWER_ENERGY_MIN' --include=*.nim .
```

Convention followed throughout: read once at module init, into a `let` with an
`EnvVar` name constant beside it, so one compiled binary can A/B every arm by
environment alone.