Aura/docs/specs/0086-cost-flag-cli-ergonomics.md

# Cost-flag CLI ergonomics — Design Spec

**Date:** 2026-06-29
**Status:** Draft — awaiting user spec review
**Authors:** orchestrator + Claude

## Goal

Close the three non-blocking ergonomics/diagnostic gaps the milestone-close
fieldtest of #148 surfaced on the `aura run` cost-flag surface
(`--cost-per-trade`, `--slip-vol-mult`, `--carry-per-cycle`), plus the one
named spec_gap. The cost-model-graph capability already ships end-to-end; this
cycle only makes its CLI surface *honest and self-describing*, changing no
cost-model node, no R chain, and no `net_r_equity` / `summarize_r` math.

Three observable defects, all instances of one anti-pattern (a misuse that
produces no actionable feedback):

1. **Units undiscoverable.** The usage string lists the three cost flags with no
   unit, scale, or per-close-vs-per-held-cycle note. A user must trial magnitudes
   to find one that moves net-R.
2. **Negative-rate diagnostic is mute.** A negative rate already exits 2, but the
   message is the bare 1-line usage dump — it never names the cause, and the
   usage never states the non-negativity constraint, so a sign typo is
   indistinguishable from a mistyped flag.
3. **Silent no-op on a non-R harness (the spec_gap).** `aura run --harness sma
   --cost-per-trade 0.001` exits 0 with output byte-identical to the no-cost run:
   the flag is silently ignored. C10 defines cost only against the R chain and
   never constrained what a cost flag does on a harness that produces no R.

## Architecture

All three changes live in `crates/aura-cli/src/main.rs`, in and around the
**pure** `parse_run_args` grammar (line 3136) — the function the codebase
deliberately keeps side-effect-free so the grammar is unit-testable; `main`
already maps its `Err(String)` to `eprintln!("aura: {msg}")` + `exit(2)` (line
3277). Keeping every new rejection inside `parse_run_args` means each is a pure,
unit-testable `Result::Err`, and the fail-loud exit path is the existing one —
no new exit site, no new I/O. One ledger note pins the decision in C10.

**Decision recorded (reference issue #153, 2026-06-29 — derived, not preference):**
item 3 resolves to **reject, exit 2** (not warn-and-continue). Basis: aura's
settled *refuse-don't-guess / fail-loud* discipline — C10 refuse-don't-guess
(exit 2 on absent geometry), the "every error path exits 2, never panics" rule
(#107), and the negative-rate guard in this very surface already exits 2. A
silent no-op is the lone exception in an otherwise fail-loud surface; rejecting
is the only behaviour consistent with every sibling guard.

## Concrete code shapes

### Worked user-facing surface (the acceptance evidence)

```text
# Item 1 — units/constraints discoverable at the point of use (--help, exit 0)
$ aura run --help
usage: aura run [--harness <sma|macd|stage1-r>] [--real <SYMBOL> [--from <ms>] [--to <ms>]] [--trace <name>] [--cost-per-trade <f64>] [--slip-vol-mult <f64>] [--carry-per-cycle <f64>] | aura chart ...
cost flags (stage1-r only; all >= 0, charged in R as cost/|entry−stop|): --cost-per-trade = per-trade cost in price units; --slip-vol-mult = slippage multiplier on realized vol; --carry-per-cycle = carry in price units per ENGINE cycle (not per day, not an overnight swap)

# Item 2 — a negative rate names the cause + flag (exit 2, stderr), no runs/ write
$ aura run --harness stage1-r --cost-per-trade -0.5
aura: --cost-per-trade must be non-negative, got -0.5
$ echo $?
2

# Item 3 — a cost flag on a non-R harness is a usage error (exit 2, stderr), no runs/ write
$ aura run --harness sma --cost-per-trade 0.001
aura: cost flags require an R-evaluator harness (stage1-r); --harness sma/macd produces no R to charge against
$ echo $?
2

# Unchanged: the accepted path is byte-identical to today
$ aura run --harness stage1-r --cost-per-trade 2          # → net_expectancy_r as before
$ aura run --harness stage1-r --carry-per-cycle 0         # → exactly free (zero-rate boundary)
```

### Implementation shape (before → after, secondary)

**(a) Non-negativity guard — named cause, factored (items 2).** The three
flags currently restate an identical guard that returns the bare usage:

```rust
// BEFORE (×3, lines 3188 / 3197 / 3206):
"--cost-per-trade" if cost.is_none() => {
    let (value, t) = t.split_first().ok_or_else(usage)?;
    let v: f64 = value.parse().map_err(|_| usage())?;
    if v < 0.0 { return Err(usage()); }      // mute: bare usage dump
    cost = Some(v);
    tail = t;
}

// AFTER: one helper, named diagnostic; the three arms call it.
fn parse_nonneg_rate(flag: &str, value: &str, usage: &impl Fn() -> String) -> Result<f64, String> {
    let v: f64 = value.parse().map_err(|_| usage())?;     // malformed → grammar usage error (unchanged)
    if v < 0.0 {
        return Err(format!("{flag} must be non-negative, got {v}"));  // named cause + offending flag
    }
    Ok(v)
}
// arm:
"--cost-per-trade" if cost.is_none() => {
    let (value, t) = t.split_first().ok_or_else(usage)?;
    cost = Some(parse_nonneg_rate("--cost-per-trade", value, &usage)?);
    tail = t;
}
```

A *malformed* value (`--cost-per-trade x`) stays a grammar usage error
(unchanged); only the *parsed-but-negative* case gets the named message.

**(b) Non-R-harness guard — post-parse, pure (item 3).** After the parse loop,
once `harness` is resolved (line 3215), before constructing `RunArgs`:

```rust
let harness = harness.unwrap_or(HarnessKind::Sma);
// Cost flags are defined only against the R chain; a non-R harness produces no R
// to charge against, so a cost flag there is a usage error (refuse-don't-guess).
if !matches!(harness, HarnessKind::Stage1R)
    && (cost.is_some() || slip_vol_mult.is_some() || carry_per_cycle.is_some())
{
    return Err(
        "cost flags require an R-evaluator harness (stage1-r); \
         --harness sma/macd produces no R to charge against".to_string(),
    );
}
```

**(c) Usage note — units + constraints, one appended line (item 1).** A
`const COST_FLAGS_NOTE: &str` appended to both the `parse_run_args` `usage`
closure (line 3137) and the top-level `USAGE` (line 3244), so the same note
shows on a grammar error and on `--help`:

```rust
const COST_FLAGS_NOTE: &str = "\ncost flags (stage1-r only; all >= 0, charged in R as cost/|entry−stop|): \
--cost-per-trade = per-trade cost in price units; --slip-vol-mult = slippage multiplier on realized vol; \
--carry-per-cycle = carry in price units per ENGINE cycle (not per day, not an overnight swap)";
```

### Ledger note (C10) — pin the spec_gap decision

Add one note to contract C10 in `docs/design/INDEX.md` (cost-model area), e.g.:

> **Realization (cost-flag harness scoping, #153).** Cost flags are defined
> only against an R-evaluator harness (the gross-R → net-R chain). On a non-R
> harness (`--harness sma`/`macd`, which produce no R) a cost flag is a **usage
> error (exit 2)**, not a silent no-op — refuse-don't-guess. Negative cost rates
> are likewise rejected (exit 2) with a named diagnostic.

## Components

- `parse_nonneg_rate` (new, pure, private) — parse + non-negativity, named error.
- `parse_run_args` (edited) — three arms call the helper; one post-parse R-harness
  guard; the `usage` closure appends `COST_FLAGS_NOTE`.
- `COST_FLAGS_NOTE` (new const) + `USAGE` (edited to append it).
- C10 ledger note (one paragraph).

## Data flow

Unchanged at runtime. `parse_run_args` either returns `Ok(RunArgs)` (the
accepted path → `run_dispatch` → identical to today) or `Err(String)` (the new
named messages → `main`'s existing `eprintln!` + `exit(2)`). No cost-model node,
harness, tap, or metric is touched.

## Error handling

Every misuse is a single, named, fail-loud stderr line + `exit(2)`, no `runs/`
write, no panic — the established CLI convention. Three messages:
malformed value → existing grammar usage; negative rate → `<flag> must be
non-negative, got <v>`; cost flag on non-R harness → `cost flags require an
R-evaluator harness (stage1-r); …`.

## Testing strategy

**Unit (pure `parse_run_args`, in `main.rs` tests):**
- `parse_run_args(["--cost-per-trade","-0.5"])` → `Err` containing `non-negative`
  and `-0.5`.
- `parse_run_args(["--carry-per-cycle","-0.001"])` and `["--slip-vol-mult","-1"]`
  → `Err` containing `non-negative` (guard symmetry across all three).
- `parse_run_args(["--harness","sma","--cost-per-trade","0.001"])` → `Err`
  containing `R-evaluator harness`.
- `parse_run_args(["--harness","macd","--slip-vol-mult","0.5"])` → `Err`.
- `parse_run_args(["--carry-per-cycle","0.5"])` (no `--harness`, defaults to sma)
  → `Err` (R-evaluator requirement).
- `parse_run_args(["--harness","stage1-r","--cost-per-trade","0.001"])` → `Ok`
  (the accepted path is unchanged).
- Malformed value still maps to the grammar usage error (regression).

**CLI integration (`crates/aura-cli/tests/cli_run.rs`):**
- `aura run --harness sma --cost-per-trade 0.001` → exit 2, stderr contains
  `R-evaluator harness`, no `runs/` write.
- `aura run --cost-per-trade -0.5` → exit 2, stderr contains `non-negative`.
- `aura run --help` → exit 0, stdout contains the cost-flags note (units +
  per-engine-cycle + `>= 0` + stage1-r-only).
- **Regression:** the existing stage1-r cost tests (cli_run.rs:1635/1688/1716/
  1738/1759/1830/1879) stay green byte-for-byte — the accepted and negative-rate
  exit-code paths are unchanged in code-behaviour; only message text and the
  previously-unguarded sma path change.

## Acceptance criteria

Judged against aura's feature-acceptance criterion (improves correctness /
removes a foot-gun; reintroduces no failure class the core constraints forbid):

- The cost-flag units, scale, per-engine-cycle semantics, non-negativity, and
  R-harness requirement are all discoverable from `aura run --help`.
- No cost-flag misuse is silent: a negative rate and a cost flag on a non-R
  harness each produce a named stderr diagnostic + exit 2, never a silent no-op
  and never a bare usage dump for these two named causes.
- The accepted stage1-r cost path and all its goldens are byte-unchanged.
- C10 carries a one-line note pinning "cost flags require an R-evaluator harness;
  non-R harness ⇒ usage error (exit 2)", closing the spec_gap.