# chassis-tune 超参数预设
#
# 一次整定的结果不只取决于被控对象与种子，还取决于"搜什么区间、用哪个场景、
# 什么算好、什么算达标、搜多细"。这份文件把这些**实验条件**固化下来：
# 评审一次、长期复用、任何机器上都能复现同一组默认行为。
#
# 用法：
#   chassis-tune presets                                        # 列出本文件全部预设
#   chassis-tune sweep --preset config/tune.toml                # 用 default_preset
#   chassis-tune sweep --preset config/tune.toml --preset-name bench_loaded
#   chassis-tune sweep --preset-name bench_loaded               # 缺省路径即本文件
#   chassis-tune sweep --preset config/tune.toml --samples 128  # CLI 覆盖预设的单项
#
# 优先级：内置默认 < 本文件的预设 < 显式命令行参数（逐字段覆盖）。
# 本文件里没写的项一律沿用内置默认，因此预设只需写"与默认不同"的部分——
# 但为了让评审者一眼看全工况，下面每个预设都写全了关键项。
#
# **安全边界**：current_limit / direction / gear_ratio / 电调 ID / 串口设备名
# 不属于超参数，本文件**不接受**这些键（写了会直接报错）。它们只来自
# chassis.toml 或显式 CLI——否则预设文件就成了绕过限流的口子。
#
# schema：见 crates/chassis-tune/src/preset.rs（version = 1，未知键报错）。

version = 1
default_preset = "sim_bench_model"

# ---------------------------------------------------------------- 离线：台架量级模型
# 默认预设。评审、CI 冒烟、改完代码回归都用它：不碰硬件，结果可复现。
[preset.sim_bench_model]
description = "离线 · 台架量级模型（静/库仑摩擦 + 2 拍延迟）：全量搜索，评审与回归的基准"

[preset.sim_bench_model.space]
# kp 覆盖手调基线 2 与仿真最优 8 两侧；对数刻度保证小增益也被采样到。
kp = { min = 0.5, max = 30.0, scale = "log" }
# ki 允许精确取 0（纯比例在某些工况已经够用）。
ki = { min = 0.1, max = 100.0, scale = "log0" }
# kd 半概率取 0：速度环加微分对量化噪声敏感，先让搜索自己决定要不要。
kd = { min = 0.01, max = 3.0, scale = "log0" }
# 前馈上限 2.5 覆盖带载所需的 ≈1.77（raw/rpm）。
feedforward = { min = 0.01, max = 2.5, scale = "log0" }

[preset.sim_bench_model.scenario]
profile = "ramp-and-hold"
target_rpm = 1500.0
ramp_ms = 200.0
hold_ms = 1800.0
dt_us = 1000.0

[preset.sim_bench_model.cost]
# 跟得上、不过冲最重要；电流只作次要惩罚（省电流不是目的，避免烧电调才是）。
settle = 1.0
overshoot = 1.0
iae = 0.5
effort = 0.1

[preset.sim_bench_model.acceptance]
max_overshoot_pct = 20.0
max_ss_error_pct = 5.0

[preset.sim_bench_model.search]
seed = 20260101
samples = 512
top_k = 8
solver = "nelder-mead"
max_iters = 200
particles = 24
max_evals = 4000
threads = 0

[preset.sim_bench_model.sim]
plant = "gm3508-bench"
# 域随机化：单点模型调出来的增益对增益/时间常数/延迟高度敏感，
# 因此默认在 8 个成员的集成上做鲁棒优化，而不是对着一条理想曲线最优化。
ensemble = 8
robust_weight = 0.5
gain_pct = 0.2
tau_pct = 0.3
delay_ticks_max = 3
friction_pct = 0.5

[preset.sim_bench_model.bench]
# 离线预设不用台架，但安全包络仍写全，方便临时切 --backend bench 预演。
max_rpm = 3000.0
max_temp_c = 70
trial_timeout_ms = 6000
between_trials_ms = 500
startup_timeout_ms = 5000
continue_on_abort = false
validate_runs = 3

# ---------------------------------------------------------------- 离线：理想模型
# 无摩擦、无延迟的一阶模型：跑得快，适合改完控制逻辑后的秒级回归，
# **不要**拿它的结果直接上车（模型里没有摩擦与延迟，最优解会偏激进）。
[preset.sim_ideal]
description = "离线 · 理想一阶模型（无摩擦无延迟）：秒级回归用，结果不可直接上车"

[preset.sim_ideal.space]
kp = { min = 0.5, max = 30.0, scale = "log" }
ki = { min = 0.1, max = 100.0, scale = "log0" }
kd = { min = 0.01, max = 3.0, scale = "log0" }
feedforward = { min = 0.01, max = 2.5, scale = "log0" }

[preset.sim_ideal.scenario]
profile = "ramp-and-hold"
target_rpm = 1500.0
ramp_ms = 200.0
hold_ms = 1800.0
dt_us = 1000.0

[preset.sim_ideal.cost]
settle = 1.0
overshoot = 1.0
iae = 0.5
effort = 0.1

[preset.sim_ideal.acceptance]
max_overshoot_pct = 20.0
max_ss_error_pct = 5.0

[preset.sim_ideal.search]
seed = 20260101
samples = 128
top_k = 4
solver = "nelder-mead"
max_iters = 100
particles = 16
max_evals = 1000
threads = 0

[preset.sim_ideal.sim]
plant = "gm3508"
# 理想模型本身就是确定的，单成员即可。
ensemble = 1
robust_weight = 0.0
gain_pct = 0.0
tau_pct = 0.0
delay_ticks_max = 0
friction_pct = 0.0

[preset.sim_ideal.bench]
max_rpm = 3000.0
max_temp_c = 70
trial_timeout_ms = 6000
between_trials_ms = 500
startup_timeout_ms = 5000
continue_on_abort = false
validate_runs = 3

# ---------------------------------------------------------------- 台架：空载
# 轮子架空（无地面载荷）。摩擦小、静差靠积分压制，前馈**必须为 0**：
# 空载加前馈会把轮子直接推到最高转速。目标转速也压低，便于观察与急停。
[preset.bench_noload]
description = "真机台架 · 空载（轮子架空）：小样本局部搜索，禁用前馈，目标 800 rpm"
# 真机后端：仍然需要 --i-know-the-wheel-spins 才会真的驱动电机。
backend = "bench"

[preset.bench_noload.space]
# 空载好解在手调基线（kp≈2, ki≈10）附近，区间收紧以提高样本密度。
kp = { min = 0.2, max = 12.0, scale = "log" }
ki = { min = 0.1, max = 50.0, scale = "log0" }
kd = { min = 0.01, max = 1.0, scale = "log0" }
# 固定为 0：单点区间 + 线性刻度 = 搜索时该轴恒定。
feedforward = { min = 0.0, max = 0.0, scale = "linear" }

[preset.bench_noload.scenario]
profile = "ramp-and-hold"
target_rpm = 800.0
ramp_ms = 300.0
hold_ms = 1700.0
dt_us = 1000.0

[preset.bench_noload.cost]
settle = 1.0
overshoot = 1.0
iae = 0.5
effort = 0.1

[preset.bench_noload.acceptance]
# 空载量化噪声相对目标更大，放宽到 25%，避免"为了 0.5% 的稳态误差过冲"。
max_overshoot_pct = 25.0
max_ss_error_pct = 8.0

[preset.bench_noload.search]
seed = 20260101
# 只有一条总线，每次试验 2s，24 次 ≈ 1 分钟：够做局部精修，不够铺满空间。
samples = 24
top_k = 3
solver = "nelder-mead"
max_iters = 40
particles = 12
max_evals = 40
threads = 1

[preset.bench_noload.sim]
# 真机后端不用这些，但保留"最接近台架的模型"，方便先离线预演再上车。
plant = "gm3508-bench"
ensemble = 8
robust_weight = 0.5
gain_pct = 0.2
tau_pct = 0.3
delay_ticks_max = 3
friction_pct = 0.5

[preset.bench_noload.bench]
# 目标是 800，过速保护留足裕量但远低于电调上限。
max_rpm = 2000.0
max_temp_c = 60
trial_timeout_ms = 6000
between_trials_ms = 500
startup_timeout_ms = 5000
continue_on_abort = false
# 同一组增益跑 3 次看重复性；离散度大说明参数贴着噪声，不该上车。
validate_runs = 3

# ---------------------------------------------------------------- 台架：带载
# 整车落地 / 配重（有地面摩擦与惯量）。静差明显，前馈开始有意义（覆盖 ≈1.77）。
[preset.bench_loaded]
description = "真机台架 · 带载（整车落地/配重）：前馈区间上调，目标 1200 rpm，超调收紧"
backend = "bench"

[preset.bench_loaded.space]
kp = { min = 0.5, max = 30.0, scale = "log" }
ki = { min = 0.5, max = 100.0, scale = "log0" }
kd = { min = 0.01, max = 3.0, scale = "log0" }
# 带载需要前馈抵消摩擦：下限不为 0，让搜索必须考虑它。
feedforward = { min = 0.5, max = 2.5, scale = "log" }

[preset.bench_loaded.scenario]
profile = "ramp-and-hold"
target_rpm = 1200.0
ramp_ms = 300.0
hold_ms = 2200.0
dt_us = 1000.0

[preset.bench_loaded.cost]
settle = 1.0
overshoot = 1.0
iae = 0.5
effort = 0.1

[preset.bench_loaded.acceptance]
# 带载下超调意味着实际位移过冲，收紧到 15%。
max_overshoot_pct = 15.0
max_ss_error_pct = 5.0

[preset.bench_loaded.search]
seed = 20260101
samples = 32
top_k = 4
solver = "nelder-mead"
max_iters = 60
particles = 16
max_evals = 80
threads = 1

[preset.bench_loaded.sim]
plant = "gm3508-bench"
ensemble = 8
robust_weight = 0.5
gain_pct = 0.2
tau_pct = 0.3
delay_ticks_max = 3
friction_pct = 0.5

[preset.bench_loaded.bench]
max_rpm = 2500.0
max_temp_c = 70
trial_timeout_ms = 8000
between_trials_ms = 800
startup_timeout_ms = 5000
continue_on_abort = false
validate_runs = 3

# ---------------------------------------------------------------- 实战验收：双向
# 上车前的最后一道关：升速 → 保持 → 降速 → 保持（回零），验的是双向跟随与
# 回零能力，比单向斜坡更接近实战。样本很少（围绕现有增益做局部复算），
# 因为此时**不该**再大改增益，只想确认现有一组在两个方向上都不出问题。
[preset.field_updown]
description = "真机实战验收 · 升降速双向（回零）：小样本复验，超调与稳态误差收紧"
backend = "bench"

[preset.field_updown.space]
kp = { min = 0.5, max = 30.0, scale = "log" }
ki = { min = 0.5, max = 100.0, scale = "log0" }
kd = { min = 0.01, max = 3.0, scale = "log0" }
feedforward = { min = 0.5, max = 2.5, scale = "log" }

[preset.field_updown.scenario]
profile = "up-down"
target_rpm = 1200.0
ramp_ms = 400.0
hold_ms = 800.0
dt_us = 1000.0

[preset.field_updown.cost]
# 双向场景里稳态误差（尤其回零）比调节时间更重要。
settle = 1.0
overshoot = 1.0
iae = 1.0
effort = 0.1

[preset.field_updown.acceptance]
max_overshoot_pct = 12.0
max_ss_error_pct = 3.0

[preset.field_updown.search]
seed = 20260101
samples = 16
top_k = 2
solver = "nelder-mead"
max_iters = 30
particles = 12
max_evals = 40
threads = 1

[preset.field_updown.sim]
plant = "gm3508-bench"
ensemble = 8
robust_weight = 0.5
gain_pct = 0.2
tau_pct = 0.3
delay_ticks_max = 3
friction_pct = 0.5

[preset.field_updown.bench]
max_rpm = 2500.0
max_temp_c = 70
# 双向场景总时长翻倍，超时也翻倍。
trial_timeout_ms = 10000
between_trials_ms = 1000
startup_timeout_ms = 5000
continue_on_abort = false
# 验收阶段多跑几次，重复性不过关就不签字。
validate_runs = 5