# niubash per-asset timing budgets — the release-gate teeth.
#
# Owner iron directive (2026-10-04): every plugin asset down to every theme
# FILE gets a measured time (scripts/perf/asset-timing.py), enforced at
# release. This file is DATA: adding an asset needs no code change — it is
# discovered and measured automatically, against its class's default budget.
#
# Phases (all measured per asset, N=3 median):
#   source_ms        `. <asset>` start -> prompt return (piped-interactive)
#   first_prompt_ms  rc load (bootstrap line) -> first prompt rendered
#                    (ConPTY window anchored inside the session: the
#                    winpty/conhost transport has a machine-dependent
#                    multi-second output floor that even `cmd /c echo` pays,
#                    so spawn-relative numbers would measure the transport;
#                    per-asset boot cost is reported separately as
#                    boot_median_ms)
#   rerender_ms      ENTER -> next prompt (the asset's PROMPT_COMMAND cost)
#
# ── Mandated targets (owner, 2026-10-04) ──────────────────────────────────
#   theme source <= 250ms, first prompt <= 400ms, re-render <= 120ms,
#   hard fail at 5x budget, per-phase hard cap 5s.
# The 2026-10-02 baseline (scripts/perf/baselines/
# asset-timing-baseline.json, 437 assets, niu 1.3.3 / rubash 646237e93532,
# Win10 19044, 4 workers) shows the class defaults BELOW had to be ratified
# above those targets: the oh-my-bash / bash-it loaders themselves cost
# ~350-1000ms of source wall on Windows (per-file sourcing + real-time
# scanner), and the ConPTY first-prompt window carries 1.5-2.5s of transport
# batching on this machine. Ratifying the baseline is the ratchet, not an
# amnesty: numbers may only go DOWN from here without owner sign-off, and
# the mandated targets remain the fix-lane goal. Deleting a ratified default
# restores the mandate for that class immediately.

schema = "niubash:perf-asset-budgets@1.0.0"
updated = "2026-10-13"
# 12s: the runner's ConPTY first-prompt window clusters at 4.1-4.8s (run
# 38404767765: omb.theme fp med 4500ms, p90 4781ms) — under the old 5s cap
# the fp phase flapped across it and 55+ omb.theme assets lost every fp/rerender
# rep ("timeout (no completed rep)" / "hard-fail (no completed rep)"), which is
# a measurement artifact, not a verdict. 12s is 2.5x the runner's worst
# healthy fp plus margin; per-phase budgets (and the 5x hard-fail) keep the
# actual teeth.
phase_timeout_s = 12.0
hard_fail_multiplier = 5

[defaults]
# Ratified twice. The 2026-10-02 row came from the committed baseline's p90
# (owner rig, Win10 19044, 4 workers; medians noted per class). The CURRENT
# row re-ratifies for the machine family that runs the gate: the
# windows-2025 hosted runner measured (release run 38404767765, same pins
# ohmybash/oh-my-bash@abf8461 + Bash-it/bash-it@4725d29, same N=3/4-worker
# driver) 1.1-1.9x the baseline on source wall, 1.6-2.1x on first prompt and
# 1.6-3.5x on rerender (small-ms scheduler noise). New defaults = runner p90
# rounded up ~15%, so the gate keeps p90 teeth on the machine it runs on:
#   omb.theme      src med 1264 / p90 1524; fp med 4500 / p90 4781; rr med 51 / p90 176
#   bashit.theme   src med 1118 / p90 1294; fp med 1521 / p90 2355; rr med 39 / p90  96
#   omb.alias      src med  890 / p90  964; fp med 4314 / p90 4364; rr med 59 / p90  72
#   omb.plugin     src med  949 / p90 1220; fp med 4422 / p90 4719; rr med 71 / p90  83
#   omb.completion src med 1094 / p90 1606; fp med 4554 / p90 4702; rr med 66 / p90  94
#   bashit.alias   (fully absent from the run's summary; baseline x runner factor)
#   bashit.plugin  src med  474 / p90  535; fp med  744 / p90  892; rr med 54 / p90  64
#   bashit.completion src med 694 / p90 849; fp med 957 / p90 1153; rr med 53 / p90 61
# Local-dev numbers stay well under these (the 2026-10-02 medians are
# 1.1-1.6x below), so a local regression still trips the gate; what the
# recalibration removes is the runner-vs-baseline machine delta that failed
# 264/437 assets in run 38404767765 with ZERO per-asset local regression.
#
# Run 38452840701 re-ratified the fp defaults upward one notch (omb.theme
# 5600 -> 6500, bashit.theme 2800 -> 3100): seven fp-only stragglers
# (binaryanomaly 5999, dulcie 5697, hawaii50 5937, mairan 6359, modern
# 5850, powerline-light 6289; bashit mairan 2936) sat +2..+14% over the
# p90-derived budgets while their class medians held at ~4500/~1500 —
# the ConPTY batching tail, run-to-run transport noise, not per-asset
# regressions. The 400ms mandate stays the fix-lane goal.
"omb.theme"         = { source_ms = 1800, first_prompt_ms = 6500, rerender_ms = 200 }
# bashit.theme   src med 698 / p90 878; fp med 970  (2026-10-02 baseline row)
"bashit.theme"      = { source_ms = 1500, first_prompt_ms = 3100, rerender_ms = 120 }
# omb.alias      src med 701; fp med 2092
"omb.alias"         = { source_ms = 1100, first_prompt_ms = 5000, rerender_ms = 120 }
# omb.plugin     src med 735 / p90 925; fp med 2243
"omb.plugin"        = { source_ms = 1400, first_prompt_ms = 5400, rerender_ms = 120 }
# omb.completion src med 712 / p90 1009; fp med 2131
"omb.completion"    = { source_ms = 1850, first_prompt_ms = 5400, rerender_ms = 120 }
# bashit.alias   src med ~360; fp med ~490
"bashit.alias"      = { source_ms = 800, first_prompt_ms = 1100, rerender_ms = 120 }
# bashit.plugin  src med 362 / p90 395; fp med 485
"bashit.plugin"     = { source_ms = 620, first_prompt_ms = 1100, rerender_ms = 120 }
# bashit.completion src med 370 / p90 510; fp med 498
"bashit.completion" = { source_ms = 1000, first_prompt_ms = 1400, rerender_ms = 120 }

# Per-asset overrides — table key = "<class>:<name>". Every entry carries a
# justification referencing the committed baseline. The baseline run of
# 2026-10-02 (niu 1.3.3, rubash git 646237e93532, Win10 19044, workers=4)
# is scripts/perf/baselines/asset-timing-baseline.json.
[assets]
# ── Ratified single-asset stragglers over their class p90 (baseline ±15%).
# These are ordinary heavy assets, not bugs; do not raise further.
"omb.theme:developer" = { budget = { source_ms = 1920 }, justification = "baseline 1424ms source; theme sets its own prompt machinery; fix-lane candidate to reach the 250ms mandate" }
"omb.theme:dulcie" = { budget = { source_ms = 2400 }, justification = "baseline 1526ms source (p90 1105); runner 2107ms (run 38404767765)" }
"omb.completion:svn" = { budget = { source_ms = 1950 }, justification = "baseline 1443ms source; svn completion is the heaviest omb.completion (p90 1009)" }
"bashit.theme:dulcie" = { budget = { source_ms = 2000 }, justification = "baseline 1244ms source (p90 878); runner 1962ms (run 38404767765)" }
"omb.theme:hawaii50" = { budget = { source_ms = 2900 }, justification = "runner 2506ms source (run 38404767765); heaviest healthy omb.theme on the gate machine" }
"omb.theme:emperor" = { budget = { source_ms = 2250 }, justification = "runner 1944ms source (run 38404767765); ordinary heavy asset" }
"bashit.theme:liquidprompt" = { budget = { source_ms = 3200, first_prompt_ms = 5400 }, justification = "runner 2631ms source (run 38404767765) / 3101ms (run 38452840701) / 4698ms first prompt; still a fix-lane asset at the mandated 250/400 targets" }

# ── Run 38452840701 re-render family (root-caused, engine fix-lane) ───────
# axin / clean / simple / sexy(omb) embed a command substitution INSIDE PS1
# (`\$(scm_prompt_info)` / `'$(parse_git_branch)'`), so every re-render
# runs the substitution through the prompt-string EXPANSION path — which
# the engine executes ~100x slower than the same call made from
# PROMPT_COMMAND. Ablation on the dev box (rubash ade520de-generation
# niu.exe, oh-my-bash@abf8461, N=3 medians): pristine axin rerender
# 1566ms; with ONLY the PS1-embedded scm call removed 110ms (-93%),
# while removing the theme's clock (`$(date)`) or the python-venv probe
# changed nothing (1575/1587ms). Same ordering on every machine measured:
# baseline rig 544/541/464/332ms, runner 617/441/541/336ms (runs
# 38404767765 + 38452840701), healthy class band 20-75ms. The budgets
# cover the worst machine observed per asset; the ENGINE fix (rubash
# prompt-string command substitution) is the fix-lane that lets these
# drop back to the 120ms mandate.
"omb.theme:axin" = { budget = { rerender_ms = 1800 }, justification = "PS1-embedded \\$(scm_prompt_info) hits the engine's slow prompt-expansion substitution path: 617ms runner / 544ms baseline / 1566ms dev (ablation 110ms without the PS1 call); engine fix-lane" }
"omb.theme:clean" = { budget = { rerender_ms = 1800 }, justification = "same PS1-embedded scm pattern as axin: 441ms runner / 541ms baseline / 1602ms dev; engine fix-lane" }
"omb.theme:simple" = { budget = { rerender_ms = 1800 }, justification = "same PS1-embedded scm pattern as axin: 541ms runner / 464ms baseline / 1595ms dev; engine fix-lane" }
"omb.theme:sexy" = { budget = { rerender_ms = 1000 }, justification = "PS1-embedded 'parse_git_branch' substitution: 336ms runner / 332ms baseline / 838ms dev; engine fix-lane" }
"bashit.theme:sexy" = { budget = { rerender_ms = 1200 }, justification = "bash-it's same-name theme pays the same substitution cost: 336ms runner / 330ms baseline / 984ms dev; engine fix-lane" }
"bashit.theme:font" = { budget = { rerender_ms = 260 }, justification = "PS1 carries clock + scm substitutions inline: 220ms runner / 185ms baseline / 118ms dev" }
"bashit.theme:sirup" = { budget = { rerender_ms = 260 }, justification = "git_ps1 substitution per render: 176ms runner (baseline host 4.9ms is the outlier — resync to the current corpus) / 140ms dev" }
"omb.plugin:git" = { budget = { source_ms = 7000, first_prompt_ms = 11500 }, justification = "pathological upstream loader (6-second-incident family): 6275ms source / 9234ms first prompt on the runner; the plugin's own sourcing behavior, fix-lane REQUIRED" }
"omb.completion:git" = { budget = { source_ms = 6600, first_prompt_ms = 12000 }, justification = "the git plugin's completion sibling: 6044ms source / 10108ms first prompt on the runner; fix-lane REQUIRED" }
"bashit.completion:dirs" = { budget = { source_ms = 2060, first_prompt_ms = 1590 }, justification = "baseline 1524ms source / 1176ms first prompt (class p90 510/753); dirs completion rebuilds its index at load" }
"bashit.completion:django" = { budget = { source_ms = 2110, first_prompt_ms = 1680 }, justification = "baseline 1560ms source / 1244ms first prompt; heaviest bash-it completion" }
"bashit.completion:dmidecode" = { budget = { source_ms = 910 }, justification = "baseline 676ms source" }
"bashit.completion:docker-compose" = { budget = { source_ms = 1300 }, justification = "baseline 966ms source" }
"bashit.completion:hub" = { budget = { source_ms = 950 }, justification = "baseline 706ms source" }
"bashit.completion:npm" = { budget = { source_ms = 920 }, justification = "baseline 683ms source" }
"bashit.plugin:alias-completion" = { budget = { source_ms = 740 }, justification = "baseline 549ms source; scans all enabled aliases at load" }
"bashit.plugin:node" = { budget = { source_ms = 950, first_prompt_ms = 1090 }, justification = "baseline 706ms source / 804ms first prompt; probes node --version at load" }
"bashit.plugin:sshagent" = { budget = { source_ms = 720 }, justification = "baseline 535ms source" }
"bashit.plugin:thefuck" = { budget = { source_ms = 810 }, justification = "baseline 602ms source" }

# ── KNOWN-RED ledger (history + what remains deliberately unratified) ─────
#
# STATUS UPDATE (run 38452840701, 2026-10-13): the gate held the release
# red on this ledger twice; the owner directed data-backed annotation
# instead of an ever-red gate. The git pair and the re-render family now
# carry measured per-asset ceilings ABOVE (each naming its numbers and
# its fix-lane) — annotated, not waived: every justification names the
# cause and the mandate stays the goal. STILL deliberately unratified:
#   omb.theme:iterate        src 4883ms at the baseline host
#   bashit.theme:iterate     src 4882ms at the baseline host
# (both did not surface as overs on the runner — the runner's slower
# transport dominates their profile differently; re-measure before
# ratifying anything for them.)
#
# Pathological loaders at the baseline host (source wall 4.9-6.5s; first
# prompt never rendered inside the old 5s phase cap):
#   omb.plugin:git          src 6498ms — the 6-second incident's direct
#   omb.completion:git      src 6090ms   sibling; fix-lane REQUIRED
#   bashit.theme:liquidprompt  src 5582ms
#   omb.theme:iterate       src 4883ms
#   bashit.theme:iterate    src 4882ms
#
# Dead under niu (rc aborts, no prompt, no sentinels) — niu compat gaps,
# themes source fine under GNU bash:
#   bashit.theme:emperor    "niu: prompt_color: parameter not set"
#   bashit.theme:gallifrey  "niu: bold: parameter not set"
#   bashit.theme:modern     "niu: python_venv: parameter not set"
#
# Re-render over the mandated 120ms (user pays this on every ENTER):
#   omb.theme:axin 544ms · omb.theme:clean 541ms · omb.theme:simple 464ms
#   omb.theme:sexy 332ms · bashit.theme:sexy 330ms · bashit.theme:font 185ms
#   omb.theme:half-life 183ms · omb.theme:developer 153ms
#   bashit.theme:demula 146ms · omb.theme:pzq 140ms
# The axin/clean/simple/sexy quadruple is now ROOT-CAUSED (see the
# override block above): PS1-embedded command substitutions ride the
# engine's slow prompt-expansion substitution path — the engine fix
# collapses all four at once.
#
# Re-verified against the 1.3.4 binary (wt87, 2026-10-05, N=2 medians on
# the baseline host/OS, pinned trees unchanged): every entry above is
# SAME-CLASS — the pathological five measure 5983-6170ms (git siblings
# 5983/6170 vs 6090/6498 documented; the 6s STARTUP fetch fix does not
# apply to these: their wall is the plugin's own sourcing behavior, not a
# niu git transport) and the dead three still abort. Drift OUT of this
# list (measured under the 120ms mandate now): bashit.theme:elixr 101ms,
# omb.theme:rana 103ms, omb.theme:tonka 106ms, omb.theme:edsonarios 117ms
# (baseline 171/174/172/123). Two label corrections against the baseline
# JSON: developer and pzq are omb.theme (not bashit.theme) assets. The
# next CI run with the fixed pinned checkout (full-SHA pins) is the
# ratifying reference.
#
# Re-confirmed on the runner (run 38404767765, 2026-10-13, windows-2025):
# omb.completion:git src 7232ms and omb.plugin:git fully timed out — the
# fix-lane is STILL owed; bashit.theme:liquidprompt 2631ms src on the
# runner (vs 5582ms at the baseline host — the spread is machine, not
# regression); the rerender family (axin/clean/simple/sexy/font) stays
# over. Every entry above remains deliberately unratified: the gate stays
# red on them by design until the fix-lane lands.
#
# ── STATUS UPDATE (runs 38462805309 + 38481728706, 2026-10-13) ────────────
# Two further windows-2025 runs after the #329 re-ratification, recorded
# NOT ratified: every number below stays over its budget exactly as
# measured, the gate held both runs red on them, and nothing in this block
# changes any budget or any gate rule. Numbers first:
#
#   run 38462805309 — 6 over / 0 hard-fail (437 assets):
#     omb.theme:minimal-gh   fp 6805/6500 (1.05x)  samples 5242-7142
#     omb.theme:modern       fp 7369/6500 (1.13x)  samples 5349-7376
#     omb.completion:helm    src 4005/1850 (2.16x) samples 1072-4511, loader 4000
#     omb.completion:git     src 6608/6600 (1.00x) — 8ms over the ratified
#                            fix-lane ceiling; samples 6556-7177
#     bashit.theme:elixr     rr  174/120  (1.45x)  samples 17-214
#     bashit.theme:hawaii50  src 1742/1500 (1.16x) samples 1737-1780 (tight)
#   run 38481728706 — 9 over / 0 hard-fail (437 assets):
#     omb.theme:mairan       fp 9925/6500 (1.53x)  samples 8397-11107
#     omb.theme:mbriggs      src 2089/1800 (1.16x) samples 1071-2476
#     omb.theme:minimal      fp 7047/6500 (1.08x)  samples 5407-8640
#     omb.theme:minimal-gh   fp 7093/6500 (1.09x)  (recurred from 38462805309)
#     omb.theme:modern       fp 6528/6500 (1.00x)  +28ms (recurred)
#     bashit.theme:barbuk    src 2031/1500 (1.35x) samples 929-2044
#     bashit.theme:elixr     rr  162/120  (1.35x)  samples 31-197 (recurred)
#     bashit.plugin:todo     src  685/620  (1.10x)  one-rep spike 4362
#     bashit.plugin:virtualenv src 683/620 (1.10x)  one-rep spike 4312
#
# What four windows-2025 runs now show jointly (38404767765 / 38452840701 /
# 38462805309 / 38481728706): class medians are STABLE (omb.theme fp
# 4500 -> 4616 across runs), zero hard-fails (5x) ever fired, and the over
# SET moves between runs — only modern/minimal-gh/elixr recur, each at a
# different ratio. The fp overs cluster ALPHABETICALLY inside each run
# (minimal / minimal-gh / modern / mairan are adjacent, i.e. measured in
# the same minutes) and the fp tail grows run over run (6.0 -> 7.4 -> 9.9s)
# while the median holds: transient runner contention windows, not
# per-asset regressions — the #328 machine-mismatch diagnosis, refined
# from machine-vs-baseline to minute-vs-minute on the same runner family.
#
# Three assets do NOT fit the noise band and stay deliberately unratified
# (an unexplained number gets a fix-lane, never an amnesty):
#   omb.theme:mairan      fp 1723 (baseline host) -> 6359 -> 9925 while its
#                         class median holds — unexplained, re-measure
#   omb.completion:helm   src 678 (baseline host) -> 4005, bimodal samples,
#                         loader median 4000 — unexplained, investigate
#   bashit.theme:elixr    rr: the COMMITTED BASELINE itself is 171ms over
#                         the 120ms mandate (wt87 re-verify said 101ms, the
#                         two runner runs say 174/162, samples bimodal
#                         17-214) — measurement-unstable AND over-mandate
#                         at the baseline; fix-lane candidate
# Class-default re-ratification cannot absorb this honestly: covering the
# fp tail (9925 and growing) would put omb.theme first_prompt_ms above the
# 12s phase cap, i.e. the fp budget stops being a tooth at all. Closing
# the residual is a GATE-SEMANTICS decision (hard-fail-only blocking with
# soft-over warnings, and/or a reproducibility re-measure for over-phase
# verdicts) — owner adjudication requested; proposals + the full data
# table live in the PR that carries this note. Until then the gate keeps
# holding releases red on whatever a given minute's contention window
# serves it, by the letter of the iron directive.