supercode-core 0.2.1

//! Acceptance tests for SPEC.md TR-4/T30 (`ReductionKind::OutputNormalized`
//! — terminal-noise normalization): ANSI/redraw collapse over bash tool
//! output, wired through the same `project`/`invert`/rehydrate machinery
//! every other reduction kind uses.
//!
//! Each test names the TR-4.md acceptance-criterion id it proves
//! (`dev/01`..`dev/05`). `crates/core/src/reduce/normalize.rs`'s own unit
//! tests cover the state-machine mechanics in isolation (SGR/CR/EL/CUU/CUD/
//! CHA, the exotic-DSR passthrough example, the truncated-escape/fuzz no-
//! panic sweep, determinism); this file covers the same reduction kind
//! wired through the real pipeline: `project_messages`, `invert`,
//! `verify_log`, `expand_reduction`, and `sidecar_search`.
//!
//! # Fixture provenance (`tests/fixtures/terminal/`)
//!
//! `cargo_build.raw` and `npm_install.raw` are GENUINE live captures taken
//! in this sandboxed worktree, byte-for-byte as emitted by the real tools —
//! not hand-typed:
//!
//! - `cargo_build.raw`: a throwaway crate (`tokio`/`clap`/`rand`/`chrono`/
//!   `itertools` deps, enough to produce a multi-crate build with cargo's
//!   real `Building [===>] N/76: crate-a, crate-b` progress bar) built under
//!   a real pseudo-terminal spawned via Python's `pty.openpty()` +
//!   `fcntl.ioctl(..., termios.TIOCSWINSZ, ...)` (cargo's progress renderer
//!   requires a real tty with a known window size — a plain pipe or the
//!   `script` command's fallback path here did not trigger it), running
//!   `cargo build --color always -j 2`, output captured verbatim from the
//!   pty master fd.
//! - `npm_install.raw`: `npm install webpack webpack-cli react react-dom
//!   typescript eslint --color=always`, captured the same way (real network
//!   access was available in this sandbox for the npm registry). Shows the
//!   real modern-npm idiom: a Braille spinner glyph + `ESC[1G` (CHA) +
//!   `ESC[0K` (EL) repeated, then the final summary lines.
//!
//! `pip_download.raw` and `docker_pull.raw` are hand-constructed,
//! deliberately modeled on the real, well-documented byte-level idioms of
//! `pip`'s `DownloadProgressBar` (`pip/_internal/cli/progress_bars.py`: a
//! `\r`-updated `|████...|` bar) and `docker pull`'s multi-layer renderer
//! (`moby/moby`'s `pkg/jsonmessage`: `ESC[<n>A` cursor-up to a layer's line,
//! `\r`, redraw, `ESC[K`, `ESC[<n>B` back down) — this sandbox has neither a
//! `pip`/`ensurepip` binary (verified: `pip`/`pip3` absent from `$PATH`,
//! `python3 -m ensurepip` reports "No module named ensurepip") nor a docker
//! daemon, so a live capture was not possible. TR-4.md's `dev/01` only
//! requires the cargo/npm fixtures to clear the 80% floor and requires
//! savings to be *reported* (not any particular figure) for pip/docker —
//! both real captures clear 80% (see
//! `cargo_and_npm_fixtures_meet_the_80_percent_reduction_floor`) and both
//! synthetic fixtures still exercise real coverage gaps the real captures
//! don't (docker's cursor-up/down multi-line redraw; pip's SGR "Successfully
//! installed" line mixed with many-frame `\r` bar updates).
//!
//! `plain_output.raw` is a plain multi-line `cargo test` result summary with
//! no `\r` and no escape sequences at all — the `dev/04` "untouched" fixture.
//!
//! Golden files (`*.golden.txt`) were generated by running this branch's own
//! `reduce::normalize::normalize` against each `.raw` fixture and are
//! checked in as goldens (a plain-text visual diff on review); they are not
//! independently authored.

use std::path::{Path, PathBuf};

use supercode::reduce::rehydrate::{expand_reduction, sidecar_search};
use supercode::reduce::{
    self, invert, normalize, project_messages, reduction_id, stub, verify_log, ReductionKind,
    ReductionLog, ReductionPolicy, REDUCTION_SENTINEL,
};
use supercode::session::Session;
use supercode::{ChatMessage, FunctionCall, Role, ToolCall};

fn fixtures_dir() -> PathBuf {
    Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/terminal")
}

fn load_raw(name: &str) -> String {
    std::fs::read_to_string(fixtures_dir().join(format!("{name}.raw")))
        .unwrap_or_else(|e| panic!("reading {name}.raw: {e}"))
}

fn load_golden(name: &str) -> String {
    std::fs::read_to_string(fixtures_dir().join(format!("{name}.golden.txt")))
        .unwrap_or_else(|e| panic!("reading {name}.golden.txt: {e}"))
}

fn pct_reduction(original: usize, normalized: usize) -> f64 {
    if original == 0 {
        0.0
    } else {
        100.0 * (1.0 - (normalized as f64 / original as f64))
    }
}

// ---- dev/01: fixture suite normalizes to final-rendered content, with the
// cargo/npm fixtures clearing an 80% byte-reduction floor ------------------

#[test]
fn cargo_and_npm_fixtures_meet_the_80_percent_reduction_floor() {
    for name in ["cargo_build", "npm_install"] {
        let raw = load_raw(name);
        let golden = load_golden(name);
        let normalized = normalize::normalize(&raw);
        assert_eq!(
            normalized, golden,
            "{name}: normalize() must match the checked-in golden"
        );
        let pct = pct_reduction(raw.len(), normalized.len());
        assert!(
            pct >= 80.0,
            "{name}: reduction {pct:.1}% must be >= 80% (orig={}, normalized={})",
            raw.len(),
            normalized.len()
        );
        println!(
            "{name}: {} B -> {} B ({pct:.1}% reduction)",
            raw.len(),
            normalized.len()
        );
    }
}

#[test]
fn pip_and_docker_fixtures_normalize_to_golden_final_rendered_content() {
    for name in ["pip_download", "docker_pull"] {
        let raw = load_raw(name);
        let golden = load_golden(name);
        let normalized = normalize::normalize(&raw);
        assert_eq!(
            normalized, golden,
            "{name}: normalize() must match the checked-in golden"
        );
        let pct = pct_reduction(raw.len(), normalized.len());
        assert!(
            pct > 0.0,
            "{name}: normalization must report a nonzero savings figure, got {pct:.1}%"
        );
        println!(
            "{name}: {} B -> {} B ({pct:.1}% reduction)",
            raw.len(),
            normalized.len()
        );
    }
}

// ---- shared harness for the full-pipeline tests below ---------------------

/// An assistant turn invoking `bash` (or another `NORMALIZE_TOOLS` name)
/// with tool-call id `call_id` — T30's candidate rule is keyed on the
/// INVOKING call's function name (paired via `tool_call_id`, exactly like
/// A8's `detect_reads`), not `ChatMessage::name` on the tool result itself
/// (see `reduce.rs`'s `detect_normalize_candidates` doc comment for why).
fn assistant_calling(call_id: &str, tool_name: &str) -> ChatMessage {
    ChatMessage {
        role: Role::Assistant,
        content: None,
        content_parts: None,
        tool_calls: Some(vec![ToolCall {
            id: call_id.to_string(),
            kind: "function".to_string(),
            function: FunctionCall {
                name: tool_name.to_string(),
                arguments: r#"{"command":"cargo build"}"#.to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

/// Build a minimal one-tool-result message slice: an assistant turn calling
/// `bash`, followed by its (noisy) tool result — enough for
/// [`project_messages`]'s T30 candidate detection (paired via
/// `tool_call_id`) to see it. `protect_last_n_tool_results` defaults to 3,
/// so a lone tool result would normally be "protected" from A7 truncation,
/// but `OutputNormalized` deliberately does not consult `protected` (see
/// `reduce.rs`'s T30 pass) — this harness intentionally exercises that most
/// recent-tool-result case, not an artificially aged one.
fn bash_result_messages(content: &str) -> Vec<ChatMessage> {
    vec![
        ChatMessage::user("please build the project"),
        assistant_calling("call-1", "bash"),
        ChatMessage::tool_result("call-1", "bash", content.to_string()),
    ]
}

fn bash_idx() -> usize {
    2
}

/// Wraps `msgs` in a bare `Session` (mirrors `reduce_dedup.rs::session_of`)
/// so it can serve as the `sidecar_session` argument `invert` needs — the
/// sidecar is just the unreduced message list.
fn session_of(msgs: Vec<ChatMessage>) -> Session {
    let mut session = Session::from_claude_code_str("").unwrap();
    session.messages = msgs;
    session
}

/// Find the (assumed unique) `OutputNormalized` reduction in `log`.
fn find_output_normalized(log: &ReductionLog) -> &reduce::Reduction {
    log.reductions
        .iter()
        .find(|r| matches!(r.kind, ReductionKind::OutputNormalized { .. }))
        .expect("expected an OutputNormalized reduction in the log")
}

// ---- dev/01 (continued) + dev/05: the reduction record + stub -------------

#[test]
fn project_messages_creates_output_normalized_reduction_for_a_noisy_bash_result() {
    let raw = load_raw("cargo_build");
    let msgs = bash_result_messages(&raw);
    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());

    assert_eq!(log.reductions.len(), 1, "{:?}", log.reductions);
    let r = find_output_normalized(&log);
    let (original_bytes, normalized_bytes) = match r.kind {
        ReductionKind::OutputNormalized {
            original_bytes,
            normalized_bytes,
        } => (original_bytes, normalized_bytes),
        _ => unreachable!(),
    };
    assert_eq!(original_bytes, raw.len());
    assert_eq!(normalized_bytes, normalize::normalize(&raw).len());
    assert!(original_bytes - normalized_bytes >= normalize::DEFAULT_MIN_SAVINGS);

    // The view's content is the normalized text, a blank-line separator,
    // then the stub placeholder — mirroring A7's own
    // `kept-prefix + "\n\n" + placeholder` construction.
    let expected_normalized = normalize::normalize(&raw);
    let view_content = view[bash_idx()].content.as_deref().unwrap();
    assert!(view_content.starts_with(&expected_normalized));
    assert!(view_content.ends_with(&r.placeholder));
    assert!(view_content.contains(REDUCTION_SENTINEL));

    // dev/05: the stub/trailer carries the reduction id via the shared `sc.
    // reduction` metadata mechanism (not just embedded in the text).
    assert_eq!(reduction_id(&view[bash_idx()]), Some(r.id.as_str()));

    // The placeholder follows the shared `[sc-reduced <kind> <id>: ...]`
    // grammar (reused rather than a bespoke sentinel — see `normalize::summary`'s
    // doc comment) so `stub::parse`/`Kind::from` work with no special-casing.
    let (kind, id, summary) = stub::parse(&r.placeholder).expect("placeholder must parse");
    assert_eq!(kind, stub::Kind::OutputNormalized);
    assert_eq!(kind.as_str(), "output-normalized");
    assert_eq!(id, r.id);
    assert!(summary.contains(&format!("{}", original_bytes)) || summary.contains("collapsed"));
}

// ---- dev/02: byte-exact invert ---------------------------------------------

fn load_codex_session() -> Session {
    Session::from_codex(
        Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/codex_session.jsonl"),
    )
    .unwrap()
}

/// Overwrite the `ordinal`-th `Role::Tool` message in `session` to look like
/// a `bash` result carrying `content` — same "overwrite an existing tool
/// message" idiom `reduce_projection.rs::pad_tool_output` /
/// `reduce_loop.rs::pad_tool_output` use for A7, extended to also set `name`
/// since T30's candidate rule is keyed on tool name.
fn bashify_tool_output(session: &mut Session, ordinal: usize, content: &str) -> usize {
    let idx = session
        .messages
        .iter()
        .enumerate()
        .filter(|(_, m)| m.role == Role::Tool)
        .map(|(i, _)| i)
        .nth(ordinal)
        .expect("session must have at least ordinal+1 tool messages");
    session.messages[idx].name = Some("bash".to_string());
    session.messages[idx].content = Some(content.to_string());
    idx
}

#[test]
fn invert_restores_raw_bytes_byte_exact_after_output_normalization() {
    let raw = load_raw("cargo_build");
    let mut session = load_codex_session();
    let idx = bashify_tool_output(&mut session, 0, &raw);

    let (view, log) = reduce::project(
        &session,
        &ReductionPolicy::default(),
        &ReductionLog::default(),
    );
    assert!(log.reductions.iter().any(
        |r| r.ptr.addr.index == idx && matches!(r.kind, ReductionKind::OutputNormalized { .. })
    ));

    verify_log(&log, &session).expect("log must verify cleanly against the sidecar");

    let restored = invert(&view, &log, &session).expect("invert must succeed");
    assert_eq!(
        restored[idx].content.as_deref(),
        Some(raw.as_str()),
        "invert must restore the exact raw (pre-normalization) bytes"
    );
    assert_eq!(
        reduction_id(&restored[idx]),
        None,
        "sc.reduction must be stripped after invert"
    );
}

#[test]
fn reduction_is_prefix_stable_across_reprojection() {
    let raw = load_raw("npm_install");
    let mut session = load_codex_session();
    let idx = bashify_tool_output(&mut session, 0, &raw);

    let (view1, log1) = reduce::project(
        &session,
        &ReductionPolicy::default(),
        &ReductionLog::default(),
    );
    // Re-projecting with the prior log as input must reproduce the SAME
    // reduction (same id, same placeholder, byte-identical content) rather
    // than minting a new one — the same prefix-stability guarantee A7/A8/A9
    // already carry.
    let (view2, log2) = reduce::project(&session, &ReductionPolicy::default(), &log1);

    assert_eq!(log1.reductions.len(), 1);
    assert_eq!(log1.reductions, log2.reductions);
    assert_eq!(view1[idx].content, view2[idx].content);
    assert_eq!(reduction_id(&view1[idx]), reduction_id(&view2[idx]));
}

// ---- dev/03: idempotence + unsupported-sequence passthrough through the
// full pipeline (not just the isolated `normalize()` unit) -----------------

#[test]
fn unsupported_escape_sequence_survives_through_the_full_pipeline() {
    // A cursor-position device-status-report (DSR, `ESC[6n`) — TR-4.md's own
    // named example of a sequence this module must never guess at — embedded
    // in an otherwise-noisy real capture, big enough to clear the savings
    // floor.
    let mut raw = load_raw("cargo_build");
    raw.push_str("before-dsr\x1b[6nafter-dsr\n");

    let msgs = bash_result_messages(&raw);
    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    let r = find_output_normalized(&log);

    let expected_normalized = normalize::normalize(&raw);
    let view_content = view[bash_idx()].content.as_deref().unwrap();
    assert!(view_content.starts_with(&expected_normalized));
    assert!(
        expected_normalized.contains("before-dsr\x1b[6nafter-dsr"),
        "the unsupported DSR sequence must survive verbatim in the normalized text: {expected_normalized:?}"
    );

    // Idempotence: normalizing the already-normalized text again changes
    // nothing further (computed independently of the view's exact
    // separator/trailer bytes, which are not this test's concern).
    assert_eq!(
        normalize::normalize(&expected_normalized),
        expected_normalized
    );

    // And it still round-trips byte-exact back to the raw pre-normalization
    // bytes (DSR sequence included) via the ordinary invert path.
    let mut session = load_codex_session();
    let idx = bashify_tool_output(&mut session, 0, &raw);
    let (view2, log2) = reduce::project(
        &session,
        &ReductionPolicy::default(),
        &ReductionLog::default(),
    );
    let restored = invert(&view2, &log2, &session).unwrap();
    assert_eq!(restored[idx].content.as_deref(), Some(raw.as_str()));
    assert_eq!(reduction_id(&view[bash_idx()]), Some(r.id.as_str()));
}

// ---- P7 safety net: a normalized rendering that is STILL oversized must
// fall through to A7's bounded truncation, never ride the wire uncapped
// (the "one reduction per message" invariant every kind in this pipeline
// shares, PLUS the wire-boundedness invariant A7 exists to guarantee) ------
//
// This test used to assert the opposite (that an oversized-after-
// normalization message is claimed by `OutputNormalized` and left
// unbounded on the wire) — that was the P7 runaway-output regression this
// fix closes: OutputNormalized only strips ANSI/CR bytes, it never bounds
// the REMAINING plain-text size, so a genuinely large, mostly-distinct
// terminal output (cargo/npm diagnostics, not just redraw noise) rode the
// wire uncapped once claimed. `reduce.rs`'s T30 pass now refuses to mint an
// `OutputNormalized` reduction whenever the normalized rendering still
// exceeds `tool_output_trigger_bytes`, leaving the message raw for A7's
// candidate scan to pick up and bound normally — reproducing this exact
// input proves the deferral, not just the "never both" invariant.
#[test]
fn oversized_normalized_output_falls_through_to_bounded_a7_truncation() {
    // Real ANSI noise (cargo_build) plus enough plain filler that the
    // NORMALIZED text still comfortably exceeds A7's default
    // `tool_output_trigger_bytes` (8,192B) — this is exactly the input that
    // used to trigger the P7 regression (OutputNormalized claiming an
    // unbounded view). Followed by 3 more (small) bash results so the noisy
    // one ages out of `protect_last_n_tool_results` (default 3) and is a
    // genuine A7 candidate on its own terms — proving the fallthrough is the
    // `reduced_this_run` bookkeeping (`reduce.rs`'s T30 pass deliberately NOT
    // claiming the index), not simply "recent tool results never get
    // truncated anyway".
    let mut raw = load_raw("cargo_build");
    raw.push_str(&"y".repeat(20_000));

    let normalized_bytes = normalize::normalize(&raw).len();
    let policy = ReductionPolicy::default();
    assert!(
        normalized_bytes > policy.tool_output_trigger_bytes,
        "test fixture must exercise the still-oversized-after-normalization case; got \
         {normalized_bytes}B normalized vs a {}B trigger",
        policy.tool_output_trigger_bytes
    );

    let mut msgs = vec![
        ChatMessage::user("please build the project"),
        assistant_calling("call-1", "bash"),
        ChatMessage::tool_result("call-1", "bash", raw.clone()),
    ];
    // A distinct tool identity ("noop", never "bash") for the padding calls:
    // this fixture predates TR-6 (T16 -- same-tool/same-args supersession),
    // and `assistant_calling` hardcodes its `command` argument regardless of
    // `tool_name` -- reusing "bash" here would incidentally make these 3
    // padding calls share TR-6's `(tool, canonicalized-args)` key with the
    // noisy `cargo build` call above, and (once aged past
    // `supersede_protect_last_n`) get it evicted as `Superseded` before ever
    // reaching this test's actual subject, T30-vs-A7 precedence.
    for i in 0..3 {
        let call_id = format!("call-{}", i + 2);
        msgs.push(assistant_calling(&call_id, "noop"));
        msgs.push(ChatMessage::tool_result(call_id, "noop", "ok"));
    }
    let noisy_idx = 2;

    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());

    let claiming: Vec<&reduce::Reduction> = log
        .reductions
        .iter()
        .filter(|r| r.ptr.addr.index == noisy_idx)
        .collect();
    assert_eq!(
        claiming.len(),
        1,
        "exactly one reduction must claim the noisy message, never two: {:?}",
        log.reductions
    );
    // The still-oversized-after-normalization message must fall through to
    // A7's bounded truncation, NOT be claimed by OutputNormalized.
    assert!(
        matches!(claiming[0].kind, ReductionKind::ToolOutputTruncated { .. }),
        "a normalized rendering still over the trigger must defer to A7's bounded truncation, \
         not be claimed by OutputNormalized (the P7 regression this fix closes): {:?}",
        claiming[0].kind
    );

    // P7 safety net, made explicit: the actual WIRE bytes for this message
    // must be bounded by `tool_output_trigger_bytes` — this is the
    // assertion the pre-fix test never made, and whose absence let the
    // regression through. A7's kept-prefix (`tool_output_keep_bytes`,
    // 4,096B) plus its placeholder trailer must land comfortably under the
    // 8,192B trigger, nowhere near the ~22KB normalized (let alone ~34KB
    // raw) size.
    let view_content = view[noisy_idx].content.as_deref().unwrap();
    assert!(
        view_content.len() <= policy.tool_output_trigger_bytes,
        "P7 safety net violated: projected wire bytes ({}) exceed tool_output_trigger_bytes \
         ({}) — a normalized-but-still-huge terminal output must never ride the wire uncapped",
        view_content.len(),
        policy.tool_output_trigger_bytes
    );
    assert!(
        view_content.len() < normalized_bytes,
        "wire content ({}) must be strictly smaller than the (still oversized) normalized \
         rendering ({normalized_bytes}) — proof A7 actually truncated raw bytes, not the \
         normalized text",
        view_content.len()
    );

    // The RAW content (never the normalized rendering) is what A7 truncated
    // from — the kept prefix is real, unnormalized bytes.
    assert!(view_content.starts_with(&raw[..policy.tool_output_keep_bytes.min(raw.len())]));
}

// ---- TR-2 vs OutputNormalized precedence: a bash output that is BOTH a
// byte-identical duplicate of an earlier tool result AND ANSI/CR-noisy must
// be claimed by TR-2's DuplicateOutput, never independently normalized —
// TR-2 runs first in `project_messages` and claims the later occurrence's
// index in `reduced_this_run` before T30/TR-4's own candidate scan ever
// looks at it. The code already gets this right (verified by inspection);
// this test closes the coverage gap by actually exercising it. -------------

#[test]
fn duplicate_ansi_noisy_output_is_deduped_not_normalized() {
    // Real ANSI/CR noise (cargo_build), well over both the TR-2 dedup floor
    // (`duplicate_output_min_bytes`, default 256B) and the normalize savings
    // floor (`terminal_output_min_savings`, default 128B) — a genuine
    // candidate for EITHER pass, repeated byte-for-byte across two separate
    // bash calls.
    let raw = load_raw("cargo_build");
    let msgs = vec![
        ChatMessage::user("please build the project twice"),
        assistant_calling("call-1", "bash"),
        ChatMessage::tool_result("call-1", "bash", raw.clone()),
        assistant_calling("call-2", "bash"),
        ChatMessage::tool_result("call-2", "bash", raw.clone()),
    ];
    let first_idx = 2;
    let second_idx = 4;

    // No protection: only two tool results exist, and the default
    // `protect_last_n_tool_results` (3) would otherwise shield the later
    // (duplicate) occurrence from ever becoming a new TR-2 candidate at
    // all, defeating the point of this test.
    let policy = ReductionPolicy {
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default()
    };
    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());

    // The later, byte-identical occurrence must be `DuplicateOutput`, never
    // `OutputNormalized` -- proof TR-2's pass claimed it before T30/TR-4's
    // own candidate scan ran.
    let second_claims: Vec<&reduce::Reduction> = log
        .reductions
        .iter()
        .filter(|r| r.ptr.addr.index == second_idx)
        .collect();
    assert_eq!(
        second_claims.len(),
        1,
        "exactly one reduction must claim the later (duplicate) occurrence: {:?}",
        log.reductions
    );
    assert!(
        matches!(second_claims[0].kind, ReductionKind::DuplicateOutput { .. }),
        "TR-2 must win precedence over OutputNormalized for a byte-identical later occurrence: \
         {:?}",
        second_claims[0].kind
    );

    // The canonical (first) occurrence, meanwhile, IS independently
    // normalized -- proof this is genuinely a TR-2-vs-normalize race on a
    // real ANSI-noisy fixture, not "normalize never fires here at all".
    let first_claim = log
        .reductions
        .iter()
        .find(|r| r.ptr.addr.index == first_idx)
        .expect("canonical occurrence must also have a reduction");
    assert!(
        matches!(first_claim.kind, ReductionKind::OutputNormalized { .. }),
        "canonical occurrence must be normalized independently: {:?}",
        first_claim.kind
    );

    // invert() restores BOTH occurrences byte-exact regardless of which
    // kind claimed which index (A6).
    let session = session_of(msgs.clone());
    let inverted = invert(&view, &log, &session).expect("invert must succeed");
    assert_eq!(inverted[first_idx].content.as_deref(), Some(raw.as_str()));
    assert_eq!(inverted[second_idx].content.as_deref(), Some(raw.as_str()));
}

// ---- dev/04: plain output (no ANSI, no CR) is left untouched --------------

#[test]
fn plain_bash_output_yields_no_reduction() {
    let raw = load_raw("plain_output");
    let msgs = bash_result_messages(&raw);
    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());

    assert!(
        log.reductions.is_empty(),
        "plain (no ANSI/CR) output must yield NO reduction: {:?}",
        log.reductions
    );
    assert_eq!(view[bash_idx()].content.as_deref(), Some(raw.as_str()));
    assert_eq!(reduction_id(&view[bash_idx()]), None);
}

// ---- dev/05: expand_reduction / sidecar_search recover the raw bytes ------

#[test]
fn expand_reduction_returns_the_exact_raw_bytes() {
    let raw = load_raw("cargo_build");
    let msgs = bash_result_messages(&raw);
    let (_view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    let r = find_output_normalized(&log);

    // `minted_view` is the messages the reductions were minted against —
    // `msgs` (unmutated `history`, in the real `Agent`), NOT the returned
    // (already-reduced) `view` (see `rehydrate.rs`'s module doc comment: the
    // reduce/project layer never mutates its input, only the derived view).
    let outcome = expand_reduction(&log, &msgs, None, &r.id, None).expect("expand must succeed");
    assert_eq!(
        outcome.content, raw,
        "expand_reduction must return the exact raw bytes"
    );
    assert_eq!(outcome.total_bytes, raw.len());
}

#[test]
fn sidecar_search_finds_raw_ansi_content_hidden_from_the_view() {
    let raw = load_raw("cargo_build");
    let msgs = bash_result_messages(&raw);
    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    let r = find_output_normalized(&log);

    // The SGR "bold bright green" sequence is all over the raw capture, and
    // is exactly what normalization strips — so a match here proves
    // `sidecar_search` is searching the hidden RAW bytes, not the (already
    // ANSI-free) normalized view. `msgs` (unmutated), not `view` — see
    // `expand_reduction_returns_the_exact_raw_bytes` above.
    let hits =
        sidecar_search(&log, &msgs, None, "\u{1b}[1m\u{1b}[32m").expect("search must succeed");
    assert!(
        !hits.matches.is_empty(),
        "expected matches in the raw ANSI bytes"
    );
    assert!(hits.matches.iter().all(|m| m.reduction_id == r.id));
    assert!(hits.matches.iter().all(|m| m.kind == "output-normalized"));

    // And the normalized VIEW itself must not contain that raw escape
    // sequence at all (proving it was truly stripped, not just hidden from
    // search).
    let view_content = view[bash_idx()].content.as_deref().unwrap();
    assert!(!view_content.contains("\u{1b}[1m\u{1b}[32m"));
}