myxce/.kiro/specs/scheduler-sim-repl-verbosity/design.md

# Design Document

## Overview
**Purpose**: This feature reduces and reorganizes the console output the `xce_scheduler --interactive` simulation REPL prints while executing `continue`, so an operator can watch a long replay's progress without wading through repetitive, low-information lines.

**Users**: XCE scheduler engineers and operators debugging scheduling behavior via `xce_scheduler -i -s <replay>`, a capability shipped by the `scheduler-core` spec.

**Impact**: Changes the text `xce_repl_on_event()` (repl_break.c) and `sim_print_decision()` (sim_loop.c) print while `sim->repl_state->mode == XCE_REPL_MODE_CONTINUE`; the REPL's command prompt string (`repl_command_loop()`, repl.c); `info worker`'s/`info resource`'s resource-value rendering and the `set` command (repl_info.c, repl_params.c); and `help`'s listing plus `break`'s unrecognized-event-type error text (repl.c, repl_break.c). No change to scheduling decisions, replay determinism, or `step` mode's per-event/per-decision detail; every other aspect of every REPL command's parsing and behavior is unchanged.

### Goals
- Suppress routine per-event echo during `continue` execution, while still surfacing structural changes (Requirement 1).
- Replace unconditional per-decision echo with one summary line per round during `continue` (Requirement 2).
- Collapse consecutive idle or unchanged rounds into a single range notification (Requirement 3).
- Compress array-job identifier lists into index ranges everywhere the REPL displays multiple job identifiers together, including `info worker`'s assigned-job listing (Requirement 4).
- Leave `step` mode's per-event/per-decision detail, and every other aspect of the REPL's existing commands, unchanged except where this feature's own requirements explicitly say otherwise (Requirement 5).
- Show the current session and timestamp in the REPL's command prompt, in both `step` and `continue` modes (Requirement 5.5).
- Render byte/megabyte-denominated resource values in a human-readable unit under operator control (`set units`), and lay out `info worker`'s/`info resource`'s listings as a grouped table (Requirement 6).
- Reorganize `help`'s command listing into labeled, consistently-aligned sections, and make `break`'s valid event types discoverable in both `help` and its error message (Requirement 7).

### Non-Goals
- New REPL commands or newly inspectable data (limit usage, pend-reason detail) — tracked separately under the `scheduler-sim-diagnostics` brief.
- Any change to scheduling/allocation logic or to the deterministic outcome of a replay.
- Non-interactive (headless, no `-i`) scheduler output.
- Resolving bug-1611 (whether `DISPATCH` decisions are structurally invisible in replay mode) — this spec only guarantees Requirement 2.4 holds regardless of that investigation's outcome (see Design Decisions, `research.md`).
- Unit-aware formatting for any resource outside the confirmed canonical-bytes list (`mem`, `tmp`, `maxmem`, `maxtmp`, `maxswap`) and the confirmed canonical-megabytes `swp` — `xce_resource_t` (`src/include/xce/types.h`) carries no units metadata at all, so an arbitrary operator-defined or plugin resource is displayed as a plain unitless number, exactly as today.

## Boundary Commitments

### This Spec Owns
- Console output formatting/verbosity logic inside `xce_repl_on_event()` (repl_break.c), the `continue`-mode gate on `sim_print_decision()`'s call site (sim_loop.c), and the new round-boundary rollup/collapse logic added at `xce_repl_post_session_check()` (repl.c).
- The new jobid-range-compression helper and every call site that adopts it, including `info worker`'s assigned-job listing (repl_info.c).
- New `xce_sched_repl_state_s` fields required to track collapse state across rounds.
- The REPL prompt string construction in `repl_command_loop()` (repl.c).
- Resource-value unit classification/scaling helpers, the `units_mode` display setting, and the table/grouping layout of `info worker`'s and `info resource`'s resource listings (repl_info.c, repl_params.c).
- `repl_print_help()`'s listing structure (repl.c) and `repl_cmd_break()`'s unrecognized-event-type error text (repl_break.c).

### Out of Boundary
- Any new `info`/`break`/`set` top-level subcommand or newly inspectable scheduler data (`set units` is a new *option* of the existing `set` command, not a new command).
- Scheduling/allocation/decision logic itself (`scheduler.c`, `src/plugins/`).
- `sim_print_decision()`'s behavior when `sim->interactive` is false (headless mode) — unchanged.
- Root-causing bug-1611 — a separate investigation; this spec only ensures its own requirements hold regardless of the answer.
- Unit-aware formatting for resources outside the confirmed bytes/megabytes list (see Non-Goals) — no generic units-metadata system is added to `xce_resource_t` or the wire protocol.

### Allowed Dependencies
- Existing `xce_sched_repl_state_s` fields (`mode`, `last_n_dispatch`/`last_n_pend`/`last_n_preempt`, `session_count`) — repl_internal.h.
- The existing `core->jobs` iteration idiom (`iter_make`/`iter_next`, `JOB_IS_PEND`) already used by `repl_info_jobs()`.
- Existing helpers `fmt_jobid_ex()` (lib/utils.c) and `repl_format_ts_human()` (repl.c).
- Must not add a new call site to `sim_loop.c`'s main loop beyond gating the two calls that already exist there (`sim_print_decision`, the `xce_repl_check_event` wrapper) — that file is documented at its line-budget ceiling (repl.h's own header comment).
- `xce_sched_repl_state_s.session_count` and `xce_core_now()`/`repl_format_ts_human()` (repl.c) for the prompt's session/timestamp content — no new state needed for that part.
- `repl_break.c`'s existing `BP_KIND_TABLE` (currently `static`/file-local) — exposed via one new shared accessor declared in `repl_internal.h`, rather than duplicated as a second hardcoded list, so `help` and the unrecognized-event-type error can never drift out of sync with each other or with the table itself.
- The existing `repl_command_loop()` call site's `readline(XCE_REPL_PROMPT)` (repl.c:439) — this feature makes the prompt argument dynamic in place; no new call site.
- `src/controller/testsuite/test.c`'s `is_bytes_valued_resource()`/`format_bytes_human()` as a design *precedent* only (same classification list, same adaptive B/KB/MB/GB/TB/PB algorithm) — not a shared symbol, since `xce_test` and `xce_scheduler` are separate binaries; the scheduler gets its own `repl_*`-named port of the same logic.

### Revalidation Triggers
- If `scheduler-sim-diagnostics` adds pend-reason data, the round-rollup collapse fingerprint (Design Decisions, `research.md`) may need to incorporate it — revisit Requirement 3.2's implementation then.
- If bug-1611 is confirmed a real dispatch-visibility gap and separately fixed, the rollup's dispatch *count* source should switch from the decision tally to the same state-derived approach already used for the pend-jobid set.
- Any change to `xce_sched_repl_state_s`'s existing `mode`/tally fields by another spec should re-check this feature's rollup logic.
- If a future spec adds real units metadata to `xce_resource_t` (see Non-Goals), the hardcoded bytes/megabytes classification lists added here should be replaced by reading that metadata instead of the name-match lists.
- If `repl_break.c`'s `BP_KIND_TABLE` gains or loses an entry (e.g. a new breakpoint-able event type), `help` and `break`'s error message pick it up automatically via the shared accessor — no separate list to update, by construction.

## Architecture

### Existing Architecture Analysis
- `repl.c`/`repl_break.c`/`repl_info.c`/`repl_params.c` form the REPL command family under `scheduler-core`. `repl.h` folds the `sim->interactive` gate into `static inline` wrapper functions so `sim_loop.c`'s own diff stays to single-line, brace-free call expressions — this feature's `sim_loop.c` change must follow that same convention rather than adding a bare `if` block to the loop body.
- `xce_sched_repl_state_s` (repl_internal.h) is the REPL's own session-scoped state struct, zero-initialized via `calloc` in `repl_state_ensure()` — new fields need no separate lifecycle.
- `xce_repl_post_session_check()` (repl.c) already runs exactly once per completed round, immediately after that round's decisions are drained, purely to check the `step`-budget pause condition today.

### Architecture Pattern & Boundary Map

**Correction (post-review)**: breakpoint-hit detection and structural-event classification both live in `repl_break.c` (`xce_repl_on_event`, `xce_repl_on_begin_round`, `xce_repl_on_decision` — each calls `repl_command_loop()` directly on a match), not in `repl.c`. Requirement 3.5 ("a structural event ends the active range") and Requirement 3.6 ("a breakpoint hit ends the active range") therefore cannot be satisfied by round-boundary logic in `repl.c` alone — `repl_break.c` needs a way to force-flush the Round Rollup's active range from its own call sites. The Round Rollup component now exposes an explicit flush entry point for this, called from both files.

```mermaid
flowchart TD
    RoundEnd[Round decisions drained] --> Query[Query core jobs for PEND state]
    Query --> Fingerprint[Build fingerprint: dispatch count, pend jobid set, preempt count]
    Fingerprint --> Compare{Matches active collapsed range}
    Compare -->|Yes| Extend[Extend active range silently]
    Compare -->|No| FlushCall[Call round rollup flush]
    FlushCall --> Begin[Begin new collapsed range with this round]
    Extend --> NextRound[Advance to next round]
    Begin --> NextRound
    StructuralEvent[Structural event classified in repl_break.c] --> FlushCall
    BreakpointHit[Breakpoint hit detected in repl_break.c] --> FlushCall
    ReplayComplete[Replay completion in repl.c] --> FlushCall
    FlushCall --> PrintRange[Print the flushed range, if one was open]
```

**Architecture Integration**:
- Selected pattern: a state accumulator attached to the existing round-boundary hook (`xce_repl_post_session_check`), exposing one shared flush entry point that `repl_break.c`'s event/breakpoint/decision hooks and `repl.c`'s replay-completion hook all call — no new hook call sites added to `sim_loop.c`.
- Domain/feature boundaries: the Round Rollup's state and flush/print logic live in `repl_rollup.c` (split out of `repl.c` post-review — see Correction below); `repl_break.c` calls the flush entry point but does not own or mutate the collapse state directly; `sim_loop.c` only gains a mode gate on its two existing calls.
- Existing patterns preserved: `repl.h`'s `sim->interactive`-gated inline-wrapper convention; `repl.c`'s shared-helper convention (`repl_format_ts_human`, `repl_split_tokens`) extended with two new shared helpers (range compression, staying in `repl.c`; and the round-rollup flush entry point, now implemented in `repl_rollup.c` per the Correction below, declared in `repl_internal.h` for `repl.c`/`repl_break.c` to call).
- New components rationale: deciding "print now vs. extend silently" requires comparing against the *previous* round's fingerprint, which must persist across separate `xce_repl_post_session_check()` invocations — hence new persistent fields on `xce_sched_repl_state_s` rather than function-local state. The flush entry point exists because three independent call sites (round boundary, breakpoint hit, replay completion) across two files all need to end an active range the same way.
- Steering compliance: one domain per file preserved; `snake_case`/`repl_*` naming conventions followed throughout. **Correction (post-review, task 2.2, supersedes "no new production `.c` file needed" above)**: task 2.2's first pass kept the Round Rollup entirely inside `repl.c`, per this bullet's original text. That pushed `repl.c` to 1160 lines — 45% over this project's 800-line-per-file cap (`src/scheduler/CLAUDE.md`) — while review confirmed the logic itself was correct. The Round Rollup's fingerprint-build/pend-jobid-query/emit/flush/round-process functions (`repl_jobid_u64_cmp` through `repl_round_rollup_process`) now live in a new `src/scheduler/src/repl_rollup.c`, following the exact domain-split precedent this same file family already established for `repl_break.c`/`repl_info.c`/`repl_params.c` (`repl_internal.h`'s own "Why a split at all" section) — one more sibling sharing state via `repl_internal.h`, not a departure from that pattern. `repl.c`'s `repl_count_jobs()` becomes non-static so `repl_rollup.c`'s pend-jobid query can call it; both it and `repl_round_rollup_process()` gain `extern` declarations in `repl_internal.h`. `repl.c`'s own wiring calls (`xce_repl_post_session_check()`/`xce_repl_on_replay_complete()`) are unchanged. Tasks 6 (REPL Prompt Context) and 9 (Help & Breakpoint Discoverability/`BP_KIND_TABLE`) remain correctly targeted at `repl.c` — different domains (prompt construction, help/breakpoint-table rendering) with no Round Rollup logic of their own, unaffected by this correction.

### Technology Stack

| Layer | Choice / Version | Role in Feature | Notes |
|-------|------------------|------------------|-------|
| Scheduler daemon (C, gnu17) | existing `src/scheduler/` toolchain, no new dependency | All logic lives in existing REPL/sim-driver translation units | Matches `tech.md`; no new library, no CGO/Go involvement |

## File Structure Plan

### Modified Files
**Correction (post-review, task 2.2)**: keeping the Round Rollup's logic inside `repl.c`, as the original `repl.c`/`repl_internal.h` entries below described, pushed that file to 1160 lines — 45% over this project's 800-line cap (`src/scheduler/CLAUDE.md`) — even though review confirmed the logic itself was correct. The Round Rollup's fingerprint/query/emit/flush/process functions now live in a new `src/scheduler/src/repl_rollup.c` instead, matching this exact file family's established domain-split precedent (`repl_break.c`/`repl_info.c`/`repl_params.c` already split out of `repl.c` by domain, per `repl_internal.h`'s own "Why a split at all" section) rather than growing `repl.c` further. The `repl.c` and `repl_internal.h` entries below, and the new `repl_rollup.c` entry, reflect this. Tasks 6 (REPL Prompt Context) and 9 (Help & Breakpoint Discoverability) are unaffected — both remain correctly targeted at `repl.c`, a different domain with no Round Rollup logic of its own.

**Accepted residual (post-review, task 2.2)**: after the split, `repl.c` settled at 808 lines — 1% over the 800-line cap, down from the original 1160-line (45% over) violation. Verified empirically (not assumed): trimming the two doc comments this task added to their practical minimum recovers only 9 of the needed 17+ lines, and deleting both entirely would still leave 801. Getting strictly under 800 would require another file split for single-digit lines, which is disproportionate churn given this codebase already tolerates comparable or larger legacy exceptions (`scheduler.c` at 1382 lines, `plugins/builtin_fcfs.c` at 1026 lines, both noted in `src/scheduler/CLAUDE.md` as "do not use... as precedent for new files" but tolerated in place). 808 lines is accepted as this task's final state; if tasks 6/9's later `repl.c` additions push it meaningfully further over, revisit with a fresh split decision at that point rather than trimming comments further now.
- `src/scheduler/src/repl_internal.h` — extend `xce_sched_repl_state_s` with round-collapse state fields (see Data Models), a new `units_mode` field (Requirement 6), and declare the new shared jobid-range-compression helper, resource-formatting helpers, and the `BP_KIND_TABLE` accessor (Requirement 7); post-review (Correction above), also declares `repl_count_jobs()` and `repl_round_rollup_process()` `extern` so `repl_rollup.c` can call the former and `repl.c` can call the latter.
- `src/scheduler/src/repl_break.c` — `xce_repl_on_event()` gains routine-vs-structural event classification gated on `rs->mode` (Requirement 1); `repl_print_event_summary()`'s `JOB_PUSH`-style cases adopt the jobid-range-compression helper (Requirement 4); `xce_repl_on_event()`, `xce_repl_on_begin_round()`, and `xce_repl_on_decision()` each call the Round Rollup's flush entry point before printing a structural-event line or before re-entering `repl_command_loop()` on a breakpoint hit (Requirement 3.5, 3.6); `repl_cmd_break()`'s unrecognized-event-type error gains the full valid-event-type listing, rendered via a new shared accessor over the existing `BP_KIND_TABLE` (Requirement 7.5); `repl_cmd_delete()` already fixed independently (bug-1761, `strtol`+endptr validation) — no further change needed here for this spec.
- `src/scheduler/src/repl.c` — `xce_repl_post_session_check()` calls the Round Rollup's round-boundary fingerprint/collapse logic, and `xce_repl_on_replay_complete()` calls its flush entry point before printing the existing replay-complete banner (Requirements 2, 3; post-review Correction above: the Round Rollup's own logic — fingerprint build, collapse, flush, emit — now lives in the new `repl_rollup.c` below, not here); implements the new shared jobid-range-compression helper (Requirement 4), following the file's existing shared-helper convention; `repl_command_loop()`'s `readline(XCE_REPL_PROMPT)` call gains a per-iteration dynamically-built prompt (Requirement 5.5); `repl_print_help()` is rewritten into labeled, consistently-aligned sections, `info config` moves next to the other `info` lines, and the `break` line grows the same `BP_KIND_TABLE`-derived event-type listing `repl_break.c`'s error message uses (Requirement 7).
- `src/scheduler/src/repl_rollup.c` (new, post-review — Correction above) — the Round Rollup's own fingerprint-build/pend-jobid-query/emit/flush/round-boundary-process logic (Requirements 2, 3), split out of `repl.c` to stay under the 800-line cap; shares state with `repl.c` via `repl_internal.h`, matching `repl_break.c`/`repl_info.c`/`repl_params.c`'s established split; added to `scheduler_srcs` in `src/scheduler/meson.build` immediately after `repl.c`.
- `src/scheduler/src/repl_info.c` — new resource-value classification/scaling helpers (byte-denominated list, `swp`'s megabyte special case, whole-number formatting) and a table/grouping rewrite of `repl_info_worker()`'s and `repl_info_resource()`'s resource-listing loops, both reading `rs->units_mode` (Requirement 6); the assigned-job listing in `repl_info_worker()` adopts the jobid-range-compression helper (Requirement 4.5).
- `src/scheduler/src/repl_params.c` — `repl_cmd_set()` gains a `units raw|human|both` branch alongside the existing `speed`/`output`/`verify` branches, storing the selection on the new `rs->units_mode` field (Requirement 6.2, 6.3).
- `src/scheduler/src/sim_loop.c` — the existing call site invoking `sim_print_decision()` gains a `continue`-mode gate, mirroring `repl.h`'s inline-wrapper convention (Requirements 2, 5).
- `src/scheduler/include/repl.h` — no functional change expected; existing inline wrappers already cover the gating points this feature needs.

### New/Modified Test Files
**Correction (post-review, task 2.1)**: `test_repl_execution_control.c` was already over this project's 800-line cap (937 lines) before this feature's work began. Landing task 2.1's Round Rollup tests there per the original plan below pushed it to 1563 lines — nearly double the cap, and past the ~1000-line threshold this exact spec family already treats as a split trigger (`test_repl_snapshot_params.c` was split out of `test_repl_breakpoints_inspection.c` for precisely this reason; see that file's own meson.build comment). Round Rollup's tests (task 2.1, and any Round-Rollup-internal portion of later tasks, e.g. Requirement 3.5's flush-on-structural-event/breakpoint interaction) now live in a new `src/scheduler/testsuite/test_repl_round_rollup.c` instead, duplicating (not sharing) whatever pre-existing static fixture helpers it needs — the established convention this same file pair (`test_repl_breakpoints_inspection.c`/`test_repl_snapshot_params.c`) already follows, rather than factoring out a shared test-helper header. `test_repl_execution_control.c` remains at its restored 937-line pre-feature size (still over cap on its own, a pre-existing condition this feature doesn't need to fix) and is the target only for the Requirement 5 no-regression checks, the new prompt-content tests (Requirement 5.5), and `help`-listing structure tests (Requirement 7.1-7.4) — re-check its line count before landing each of those; if it approaches ~1000+ lines again, split further rather than growing past it.

**Correction (post-review, task 3.1)**: applying the same reasoning that split Round Rollup's tests out in task 2.1 — Requirement 1's Event Echo Gate tests are comparable in scope (4 behavioral cases: routine-event suppression, step-mode-unaffected, structural-event-ends-range, breakpoint-ends-range) and would push the already-over-cap `test_repl_execution_control.c` (937 lines) well past 1300+ lines. They instead live in a new `src/scheduler/testsuite/test_repl_event_echo.c`, duplicating pre-existing fixture helpers per this file family's established convention, decided proactively before implementation rather than after a review round.
- `src/scheduler/testsuite/test_repl_execution_control.c` — the Requirement 5 no-regression checks (this file already drives `sched_sim_driver_run()` end-to-end and captures real process output via `child_continue_then_inspect`-style helpers); also gains the new prompt-content tests (Requirement 5.5) and `help`-listing structure tests (Requirement 7.1-7.4) — the natural home given it already exercises `readline()`/prompt-adjacent behavior and no `help` test exists anywhere today.
- `src/scheduler/testsuite/test_repl_round_rollup.c` (new, per the task 2.1 Correction above) — child-process output-capture tests for Requirements 2 and 3 (per-round decision summary, collapsing, the shared flush entry point, and its call from `xce_repl_on_replay_complete`).
- `src/scheduler/testsuite/test_repl_event_echo.c` (new, per the task 3.1 Correction above) — child-process output-capture tests for Requirement 1 (mode-aware event echo classification, ending an active collapsed range on structural events/breakpoint hits).
- `src/scheduler/testsuite/test_repl_output_compression.c` (new) — focused unit tests for the jobid-range-compression helper (Requirement 4): a pure function, tested without the process harness.
- `src/scheduler/testsuite/test_repl_resource_formatting.c` (new) — focused unit tests for the byte/megabyte classification and adaptive human-scaling helpers (Requirement 6.1, 6.2, 6.5): pure functions, tested without the process harness, mirroring `test_repl_output_compression.c`'s pattern.
- `src/scheduler/testsuite/test_repl_breakpoints_inspection.c` — extended with `info worker`/`info resource` table-layout and units-mode-display tests (Requirement 6.6, alongside its existing `info worker`/`info resource` tests) and `break`'s enriched unrecognized-event-type error test (Requirement 7.5, alongside the existing `test_break_unrecognized_event_type_rejected`).
- `src/scheduler/testsuite/test_repl_snapshot_params.c` — extended with `set units raw|human|both` parsing/validation tests (Requirement 6.2, 6.3), alongside its existing `set speed`/`set output`/`set verify` tests.

## System Flows
See the Architecture Pattern & Boundary Map diagram above — it is the only non-trivial flow this feature introduces (a per-round branching decision with two exit points: silent extension, or flush-and-restart). Per-event classification (Requirement 1) is a simple, non-branching gate and does not warrant a separate diagram. The three later-added components (REPL Prompt Context, Resource Value Formatter, Help & Breakpoint Discoverability) are likewise simple, single-purpose lookups/formatters with no multi-step state machine of their own — a name-list check, an adaptive-unit loop, and a table lookup, respectively — so none of them warrant a separate diagram either.

## Requirements Traceability

| Requirement | Summary | Components | Interfaces | Flows |
|-------------|---------|-------------|-----------|-------|
| 1.1, 1.2, 1.3, 1.4 | Mode-aware event echo | Event Echo Gate | `xce_repl_on_event` | n/a (per-event, upstream of round boundary) |
| 2.1, 2.2, 2.3, 2.4 | Per-round decision summary | Round Rollup | `xce_repl_post_session_check`, state-derived pend query | Round-boundary flow |
| 3.1, 3.2, 3.3, 3.4 | Collapsing repetitive rounds (idle/unchanged detection, range content) | Round Rollup (collapse state) | fingerprint compare / flush | Round-boundary flow |
| 3.5, 3.6 | Ending an active collapsed range on outcome change, structural event, breakpoint hit, or replay completion | Round Rollup (flush entry point, and its own call from `xce_repl_on_replay_complete`), Event Echo Gate (calls it on structural events/breakpoint hits) | `repl_round_collapse_flush` | Round-boundary flow |
| 4.1, 4.2, 4.3, 4.4, 4.5 | Job identifier range compression | Jobid Range Compressor | `repl_fmt_jobid_range` | n/a (pure formatting) |
| 5.1, 5.2, 5.3, 5.4 | No regression to step mode / other commands, outside this feature's explicit exceptions | Event Echo Gate, `sim_print_decision` gate | mode checks | n/a |
| 5.5 | REPL prompt shows session/timestamp | REPL Prompt Context | `repl_build_prompt` | n/a (per-iteration, upstream of dispatch) |
| 6.1, 6.2, 6.3, 6.4, 6.5 | Human-readable resource value formatting, `set units` option | Resource Value Formatter | `repl_is_bytes_valued_resource`, `repl_format_bytes_human`, `repl_cmd_set`'s `units` branch | n/a (pure formatting + state) |
| 6.6 | `info worker`/`info resource` table/grouping layout | Resource Value Formatter | `repl_info_worker`, `repl_info_resource` | n/a (rendering only) |
| 7.1, 7.2, 7.3 | `help` listing reorganized into labeled, aligned sections | Help & Breakpoint Discoverability | `repl_print_help` | n/a (rendering only) |
| 7.4, 7.5, 7.6 | `break` event types discoverable in `help` and its error message; no other command's existence, parsing, or output changes | Help & Breakpoint Discoverability | `repl_bp_kind_names` (shared accessor over `BP_KIND_TABLE`) | n/a (pure formatting) |

## Components and Interfaces

| Component | Domain/Layer | Intent | Req Coverage | Key Dependencies (P0/P1) | Contracts |
|-----------|---------------|--------|---------------|---------------------------|-----------|
| Event Echo Gate | REPL / repl_break.c | Classify each event as routine or structural and gate the per-event echo by mode | 1.1-1.4 | `rs->mode` (P0), Round Rollup flush entry point (P0) | State |
| Round Rollup | REPL / repl_rollup.c (post-review; see Correction, was repl.c) | Build a per-round fingerprint, print it, or extend/flush a collapsed range; expose a flush entry point callable from breakpoint/structural-event/replay-complete sites | 2.1-2.4, 3.1-3.6 | `core->jobs` query (P0), existing decision tally (P0) | State, Service |
| Jobid Range Compressor | REPL / repl.c (shared helper) | Compress array-job identifier lists into index ranges | 4.1-4.5 | `fmt_jobid_ex` decomposition (P0) | Service |
| `sim_print_decision` gate | Sim driver / sim_loop.c | Suppress the direct per-decision print when the Round Rollup will report it instead | 2.3, 5.1 | REPL mode state (P0) | State |
| REPL Prompt Context | REPL / repl.c | Build the `readline()` prompt string from the current session and timestamp, falling back to the plain prompt before any session begins | 5.5 | `rs->session_count` (P0), `xce_core_now`/`repl_format_ts_human` (P0) | Service |
| Resource Value Formatter | REPL / repl_info.c, repl_params.c | Classify a resource's canonical wire unit (bytes/megabytes/unitless), render it per the operator's `units_mode`, and lay out `info worker`/`info resource`'s listings as a grouped table | 6.1-6.6 | `is_bytes_valued_resource`/`format_bytes_human` (test.c) as design precedent (P0), `rs->units_mode` (P0) | Service, State |
| Help & Breakpoint Discoverability | REPL / repl.c, repl_break.c | Group and align `help`'s listing, and expose `BP_KIND_TABLE`'s full valid-event-type set to both `help` and `break`'s error message via one shared accessor | 7.1-7.6 | `BP_KIND_TABLE` (P0) | Service |

### REPL Output (repl.c / repl_break.c / repl_info.c / repl_params.c / sim_loop.c)

#### Event Echo Gate
| Field | Detail |
|-------|--------|
| Intent | Classify each decoded event and suppress its individual echo line under `continue`, while breakpoint matching still evaluates every event |
| Requirements | 1.1, 1.2, 1.3, 1.4 |

**Responsibilities & Constraints**
- Structural events (always echoed, any mode): `CONFIG_PUSH`, `JOBS_PUSH`, `WORKERS_PUSH`, `RESOURCES_PUSH`, `RESVALS_FULL` — population/config arrivals an operator needs to see even while `continue` is otherwise quiet. Before printing a structural event's line, this gate calls the Round Rollup's flush entry point so any active collapsed range is closed and reported first (Requirement 3.5) — a structural event and a stale collapsed range must never interleave out of order.
- Routine events (echoed only in `step` mode): `RESVALS_DELTA`, `JOB_STATE_CHANGE`, `JOB_UPDATE`, `QUEUE_UPDATE`, `JOB_RUSAGE`. `JOB_STATE_CHANGE` is classified routine deliberately — the Round Rollup (Requirement 2.4) already surfaces job status transitions at round granularity, so an individual `JOB_STATE_CHANGE` line under `continue` would duplicate that information.
- Breakpoint matching (`repl_find_event_breakpoint`, plus the equivalent checks in `xce_repl_on_begin_round`/`xce_repl_on_decision`) runs on every event regardless of echo suppression — Requirement 1.4 is unaffected by this gate. On a match, before re-entering `repl_command_loop()`, the matching hook calls the Round Rollup's flush entry point so the operator sees the collapsed range's final state before the prompt returns (Requirement 3.6).

**Accepted residual (post-review, task 3.2)**: implementing this gate (classifier + 3 flush call sites + `JOBS_PUSH`'s range-compression adoption) grew `repl_break.c` from 827 to 905 lines (13% over the 800-line cap; the file was already 827/3.4% over before this task, a pre-existing condition unrelated to this spec). Unlike Round Rollup's task 2.2, this logic is wiring inside `repl_break.c`'s own pre-existing driver-facing hooks (`xce_repl_on_event`/`xce_repl_on_begin_round`/`xce_repl_on_decision`/`repl_print_event_summary`) rather than a relocatable standalone subsystem — there is no cohesive block to extract into a sibling file without fragmenting those functions. Accepted in place, same reasoning as `repl.c`'s task-2.2 residual above.

**Dependencies**
- Inbound: `xce_repl_on_event` — calls this gate before its existing breakpoint check (P0)
- Outbound: `repl_print_event_summary` — invoked only when the gate allows it (P0); Round Rollup's flush entry point — called before a structural-event line or a breakpoint-triggered command-loop re-entry (P0)

**Contracts**: Service [ ] / API [ ] / Event [ ] / Batch [ ] / State [x]

##### State Management
- State model: reads `sim->repl_state->mode` (existing field); no new state owned by this component.
- Persistence & consistency: none — a per-call decision.
- Concurrency strategy: none required — single-threaded sim driver, same as every other REPL hook.

#### Round Rollup
| Field | Detail |
|-------|--------|
| Intent | Once per completed round in `continue` mode, compute a fingerprint and either extend an active collapsed range or flush it and begin reporting the new state |
| Requirements | 2.1, 2.2, 2.3, 2.4, 3.1, 3.2, 3.3, 3.4, 3.5, 3.6 |
| File | `src/scheduler/src/repl_rollup.c` (post-review, task 2.2 — was `repl.c`; see the Architecture Integration / Modified Files Corrections above for why) |

**Responsibilities & Constraints**
- Active only while `sim->interactive && rs->mode == XCE_REPL_MODE_CONTINUE`; `step` mode keeps today's full per-decision path unchanged (Requirement 2.3, 5.1).
- Pend-jobid set is sourced by directly querying `core->jobs` for `JOB_IS_PEND` state (same iteration idiom as `repl_info_jobs()`) — not solely from the decision tally, so Requirement 2.4 holds independent of bug-1611 (see `research.md` Design Decisions).
- Dispatch/preempt counts are sourced from the existing `last_n_dispatch`/`last_n_preempt` tally.
- Must flush (print) any active collapsed range before: a breakpoint-triggered re-entry into `repl_command_loop()`, replay completion, or a fingerprint change from the active range (Requirement 3.5, 3.6). Because the breakpoint and structural-event triggers originate in `repl_break.c`, not `repl.c`, this component exposes a flush entry point other files call rather than owning every trigger site itself.

**Dependencies**
- Inbound: `xce_repl_post_session_check` (repl.c) — the round-boundary call site the fingerprint/collapse logic attaches to, calling `repl_rollup.c`'s `repl_round_rollup_process()` (P0); the Event Echo Gate (`xce_repl_on_event`) and the breakpoint checks in `xce_repl_on_begin_round`/`xce_repl_on_decision` (all repl_break.c) — call the flush entry point (P0); `xce_repl_on_replay_complete` (repl.c) — calls the flush entry point before its existing banner (P0). All three inbound call sites live in a different file from this component's own implementation (post-review, task 2.2 — see Correction above) and reach it via `repl_internal.h`'s `extern` declarations.
- Outbound: `core->jobs` iterator (P0), `repl.c`'s `repl_count_jobs()` (P0, `extern`-declared in `repl_internal.h` post-review for this cross-file call), `xce_repl_on_decision`'s existing tally fields (P0), Jobid Range Compressor (repl.c) for rendering the pend-jobid list (P0)

**Contracts**: Service [x] / API [ ] / Event [ ] / Batch [ ] / State [x]

##### Service Interface
```c
/*
 * Ends the active collapsed round range, if one is open, printing its
 * summary line immediately. Safe to call when no range is active (a
 * no-op). Called from the round boundary on a fingerprint change, and
 * from repl_break.c before a structural-event line or a breakpoint-
 * triggered command-loop re-entry, and from xce_repl_on_replay_complete
 * before its existing banner.
 */
void repl_round_collapse_flush(xce_sched_sim_driver_t* sim);
```
- Preconditions: `sim`/`sim->repl_state` non-NULL (same "safe no-op on NULL" discipline every other REPL hook in this file family already follows).
- Postconditions: no collapsed range is active after the call; if one was active, its summary line has been printed exactly once.
- Invariants: idempotent — calling it again immediately afterward with no intervening round is a no-op, never a duplicate print.

##### State Management
- State model: new fields on `xce_sched_repl_state_s` hold the active collapsed range (see Data Models) — whether one is open, its round/timestamp bounds, and its cached fingerprint.
- Persistence & consistency: in-memory, scoped to one REPL session; zero-initialized for free by `repl_state_ensure()`'s existing `calloc`.
- Concurrency strategy: none required — single-threaded sim driver.

#### Jobid Range Compressor
| Field | Detail |
|-------|--------|
| Intent | Group a list of jobids by array base and render contiguous index runs as compressed ranges |
| Requirements | 4.1, 4.2, 4.3, 4.4, 4.5 |

**Responsibilities & Constraints**
- Pure, read-only formatting function — never mutates its input list or any scheduler/job state.
- Shared by `repl_print_event_summary()`'s job-arrival listings, the Round Rollup's pend-jobid display, and `repl_info_worker()`'s assigned-job listing (Requirement 4.5) — the same helper, no per-caller variant.

**Dependencies**
- Inbound: `repl_print_event_summary` (P0), Round Rollup (P0), `repl_info_worker` (repl_info.c) (P0)
- Outbound: `fmt_jobid_ex` (lib/utils.c) — reuses its existing `(cluster, apiserver, timestamp, job[, idx])` decomposition (P0)

**Contracts**: Service [x] / API [ ] / Event [ ] / Batch [ ] / State [ ]

##### Service Interface
```c
/*
 * Formats `njobids` raw jobids into `buf` (bufsz-bounded): groups by
 * array base (cluster, apiserver, timestamp, job) and compresses
 * contiguous [idx] runs into [lo-hi] form, comma-joining non-contiguous
 * runs within one base; non-array jobids are listed individually.
 * Returns the number of bytes written (excluding the NUL terminator),
 * or a negative value if buf was too small to hold the full listing.
 */
int repl_fmt_jobid_range(char* buf, size_t bufsz,
                          const uint64_t* jobids, int njobids);
```
- Preconditions: `jobids` non-NULL whenever `njobids > 0`; `buf`/`bufsz` describe a valid, writable buffer.
- Postconditions: `buf` holds a NUL-terminated, comma-separated listing; distinct array bases appear in input order, indices ascend within each base's compressed ranges.
- Invariants: never fabricates or silently drops a jobid — the compressed listing always represents exactly the same id set the uncompressed listing would show.

#### `sim_print_decision` Gate
| Field | Detail |
|-------|--------|
| Intent | Suppress `sim_print_decision()`'s direct per-decision stdout line when the Round Rollup will report the round instead |
| Requirements | 2.3, 5.1 |

**Responsibilities & Constraints**
- Skip the direct print only when `sim->interactive && sim->repl_state->mode == XCE_REPL_MODE_CONTINUE`; `step` mode and non-interactive (headless) runs are unaffected (Requirement 5.1, 5.4).
- Implemented as an inline-wrapper-style gate at the existing call site in `sched_sim_driver_run()`, consistent with `repl.h`'s established convention, rather than a bare `if` block added to the loop body.

**Contracts**: Service [ ] / API [ ] / Event [ ] / Batch [ ] / State [x]

##### State Management
- State model: reads `sim->interactive` and `sim->repl_state->mode` (both existing fields); owns no new state.

#### REPL Prompt Context
| Field | Detail |
|-------|--------|
| Intent | Build the string passed to `readline()` each loop iteration so it shows the current session and timestamp instead of the static `"(sim) "` prompt |
| Requirements | 5.5 |

**Responsibilities & Constraints**
- Applies in both `step` and `continue` modes — it changes the prompt, not any command's output, so it sits outside Requirement 5's "leave unchanged" boundary by design (Requirement 5.5 is one of that requirement's own named exceptions).
- Before any session has begun (`rs->session_count == 0`, the same zero-initialized value `repl_state_ensure()`'s `calloc` already guarantees), falls back to the original plain `"(sim) "` prompt — there is no session/timestamp to show yet, and a fabricated placeholder value would be misleading.
- Once at least one session has begun, renders `"(sim session=<N> t=<human-ts>) "` using `rs->session_count` and `xce_core_now(&sim->core)` formatted through the existing `repl_format_ts_human()` — no new timestamp source, no raw-epoch-nanosecond component (unlike the four existing `t=%llu (...)` print sites elsewhere in this file family, which are out of this feature's scope per Boundary Commitments). `repl_format_ts_human()`'s existing output is the full ISO8601 form (`strftime`'s `"%Y-%m-%dT%H:%M:%SZ"`, e.g. `2026-09-02T03:29:38Z`), not a time-only string — the prompt reuses that format exactly rather than inventing a shorter one, since Allowed Dependencies commits to no new timestamp source.
- Uses **bare** `rs->session_count`, matching `info session`'s/`info config`'s existing convention (`repl_info.c:848,898`) — **not** `session_count + 1`. That `+1` exists only in `repl_print_event_summary()` (repl_break.c:611) to label events echoed *mid-stream*, before the session containing them has finished and incremented the count; the prompt, by contrast, is only ever shown *after* a `step`/`continue` has returned control (i.e. after `xce_repl_on_begin_round()` for the just-finished session has already run), so the bare count is the session that just completed — the same number `info session` would report if queried at that exact prompt. Copying the `+1` convention here would make the prompt disagree with `info session` by one.
- Built fresh on every `repl_command_loop()` iteration, immediately before the `readline()` call it feeds — never cached across iterations, since the session/timestamp can change between prompts.

**Dependencies**
- Inbound: `repl_command_loop()` (repl.c) — calls this immediately before its existing `readline(XCE_REPL_PROMPT)` line
- Outbound: `repl_format_ts_human()` (repl.c, existing), `xce_core_now()` (core_clock.c, existing)

**Contracts**: Service [x] / API [ ] / Event [ ] / Batch [ ] / State [ ]

##### Service Interface
```c
/*
 * Builds this iteration's REPL prompt into `buf` (bufsz-bounded):
 * "(sim) " if no session has begun yet (rs->session_count == 0), else
 * "(sim session=<N> t=<ISO8601-ts>) " using the bare current session
 * count (NOT +1 -- see repl_print_event_summary()'s different mid-stream
 * convention) and xce_core_now()/repl_format_ts_human() (same
 * "%Y-%m-%dT%H:%M:%SZ" form used elsewhere, not a shortened one). Called
 * fresh from repl_command_loop() before every readline() -- never cached.
 */
void repl_build_prompt(xce_sched_sim_driver_t* sim, char* buf, size_t bufsz);
```
- Preconditions: `sim`/`sim->repl_state` non-NULL (same discipline every other REPL hook in this file family follows); `buf`/`bufsz` describe a valid, writable buffer sized for the longest possible rendering (session count width + fixed-width human timestamp).
- Postconditions: `buf` holds a NUL-terminated prompt string; identical across repeated calls within the same session/timestamp (a pure function of `sim`'s current state).
- Invariants: never blocks, never reads input, never mutates `sim`/`sim->repl_state` — a read-only formatter, same class as the Jobid Range Compressor.

#### Resource Value Formatter
| Field | Detail |
|-------|--------|
| Intent | Classify a resource's canonical wire unit, render its value per the operator's selected display mode, and lay out `info worker`'s/`info resource`'s listings as a grouped table instead of a flat line-per-resource dump |
| Requirements | 6.1, 6.2, 6.3, 6.4, 6.5, 6.6 |

**Responsibilities & Constraints**
- Classification is a small, explicit, hardcoded name list — `repl_is_bytes_valued_resource(name)` returns true for exactly `mem`, `tmp`, `maxmem`, `maxtmp`, `maxswap` (confirmed canonical-bytes on the wire: `node_information.c`'s `node_info_get_mem()`/`node_info_get_tmp()` docstrings, `xcecd.c`'s `init_host_attrs()` bug-1554 comment). This mirrors `src/controller/testsuite/test.c`'s `is_bytes_valued_resource()` exactly (same 5 names, same rationale: "a small, explicit, client-side list rather than a wire field").
- `swp` is a confirmed special case, deliberately **not** added to the bytes list: `node_info_get_swap()`'s own docstring states its canonical wire unit is megabytes, not bytes (unlike its `maxswap` counterpart, which the bug-1554 comment confirms IS bytes). Before formatting, `swp`'s value is converted MB→bytes (`× 1024.0 × 1024.0`) and then passed through the same human-scaling formatter as the bytes-native resources, so it renders on the same adaptive B/KB/MB/GB/TB/PB scale rather than staying a raw MB float.
- `repl_format_bytes_human(bytes, buf, bufsz)` adaptively scales to the smallest unit the value fits (B/KB/MB/GB/TB/PB), mirroring `test.c`'s `format_bytes_human()` algorithm exactly (divide by 1024 while `>= 1024.0`, cap at PB).
- Display mode (`raw`/`human`/`both`) is a new `xce_sched_repl_state_s` field, `units_mode`, defaulting to `both` (Requirement 6.3) and settable via `set units raw|human|both` (repl_params.c, alongside the existing `speed`/`output`/`verify` branches) — `both` renders `"<raw> (<human>)"`, matching `xce_test`'s own `cmd_resval` display convention exactly (`test.c`'s `"%f (%s)\n"`).
- Numeric resources with no canonical unit (load averages, factors, counts — everything not in the bytes list and not `swp`) print without a unit suffix, and without a trailing `.00` when the value has no fractional part — replaces the current unconditional `"%.2f"` in `repl_info.c`'s `Value_u_fvalue` branch with a conditional format.
- `info worker`'s and `info resource`'s resource-listing loops are rewritten from an unstructured per-resource `fprintf` into an aligned table, grouped into labeled sections where applicable to that listing (consumable/capacity resources with a `used=` split; load/utilization metrics; static hardware attributes such as `type`/`model`/`cpuf`/`maxcpus`/`ndisks` — `info resource`'s cluster-wide GLOBAL/SHARED scope, per bug-1593, may not populate every section). Column widths are computed once (a shared field-width format), not hand-padded per line.
- Never alters the underlying stored/wire value or the deterministic outcome of a replay (Requirement 6.4) — display-only, same discipline `test.c`'s own `format_bytes_human()` doc comment states ("the stored/wire value stays raw bytes always, this never feeds back into anything").

**Dependencies**
- Inbound: `repl_info_worker()`, `repl_info_resource()` (repl_info.c) (P0); `repl_cmd_set()` (repl_params.c) (P0)
- Outbound: none beyond standard C formatting — a self-contained classification/formatting pair, same shape as the Jobid Range Compressor

**Contracts**: Service [x] / API [ ] / Event [ ] / Batch [ ] / State [x]

##### Service Interface
```c
/* Returns true iff `name`'s canonical wire unit is bytes (mem, tmp,
 * maxmem, maxtmp, maxswap only -- swp is handled separately, see above). */
bool repl_is_bytes_valued_resource(const char* name);

/* Adaptively scales `bytes` to the smallest unit it fits (B/KB/MB/GB/TB/
 * PB), writing e.g. "20.56 GB" into `buf` (bufsz-bounded). Display-only;
 * never touches the underlying stored value. */
void repl_format_bytes_human(double bytes, char* buf, size_t bufsz);
```
- Preconditions: `name` non-NULL for the classifier; `buf`/`bufsz` describe a valid, writable buffer for the formatter.
- Postconditions: the classifier returns a stable, name-only-dependent boolean; the formatter writes a NUL-terminated, human-readable string.
- Invariants: both are pure functions — no side effects, no dependency on REPL session state.

##### State Management
- State model: new `units_mode` field (enum `raw`/`human`/`both`) on `xce_sched_repl_state_s`, defaulting to `both`; read by the table-rendering code, written only by `set units`.
- Persistence & consistency: in-memory, scoped to one REPL session. Unless the `both` enumerator is deliberately assigned value 0 in its `enum` declaration, `calloc`'s zero-initialization will *not* select it — `repl_state_ensure()` must explicitly set `units_mode = XCE_REPL_UNITS_BOTH` after allocation regardless of member order, the one field in this struct that cannot rely on zero-initialization alone (called out explicitly since every other field in this struct today deliberately relies on `calloc`'s zero value being the correct default).
- Concurrency strategy: none required — single-threaded sim driver.

#### Help & Breakpoint Discoverability
| Field | Detail |
|-------|--------|
| Intent | Reorganize `help`'s listing into labeled, aligned sections, and expose `break`'s complete valid-event-type set to both `help` and its own error message from one shared source |
| Requirements | 7.1, 7.2, 7.3, 7.4, 7.5, 7.6 |

**Responsibilities & Constraints**
- `repl_print_help()` (repl.c) is rewritten into four labeled sections — Execution Control (`step`/`continue`/`quit`/`help`), Breakpoints (`break`/`info break`/`delete`), State Inspection (`info session`/`job`/`jobs`/`queue`/`worker`/`resource`/`summary`/`config` — `info config` moved here, next to its siblings, per Requirement 7.3), Snapshot & Parameters (`snapshot`/`set speed`/`set output`/`set verify`/`set units`) — each section using one shared field-width `fprintf` format (e.g. `"  %-38s%s\n"`) instead of independently hand-counted spaces per line, so a future command addition can't silently break alignment the way `info config`'s append did.
- `BP_KIND_TABLE` (repl_break.c, `static`/file-local today) is exposed through one new shared accessor, `repl_bp_kind_names()`, that renders the complete comma-joined list of valid `<EVENT_TYPE>` tokens once. Both `repl_print_help()`'s `break` line (repl.c) and `repl_cmd_break()`'s unrecognized-event-type error (repl_break.c, same file as the table — a trivial local call) use this accessor rather than each hardcoding the list — the two can never drift out of sync with each other or with `BP_KIND_TABLE` itself (Revalidation Triggers).
- No change to which commands exist, how any command parses its arguments, or any other command's output (Requirement 7.6) — this component only touches `help`'s rendering and the two places that now surface `BP_KIND_TABLE`'s contents.

**Dependencies**
- Inbound: `repl_print_help()` (repl.c) (P0), `repl_cmd_break()`'s error path (repl_break.c) (P0)
- Outbound: `BP_KIND_TABLE` (repl_break.c, existing) (P0)

**Contracts**: Service [x] / API [ ] / Event [ ] / Batch [ ] / State [ ]

##### Service Interface
```c
/*
 * Renders BP_KIND_TABLE's complete set of valid break <EVENT_TYPE> tokens
 * as one comma-joined, NUL-terminated string into `buf` (bufsz-bounded) --
 * e.g. "CONFIG_PUSH, JOB_PUSH, ..., DECISION". Used by both
 * repl_print_help() and repl_cmd_break()'s unrecognized-event-type error
 * so the two can never disagree with each other or with BP_KIND_TABLE.
 */
void repl_bp_kind_names(char* buf, size_t bufsz);
```
- Preconditions: `buf`/`bufsz` describe a valid, writable buffer sized for the full table (12 entries at the time of writing).
- Postconditions: `buf` holds every `BP_KIND_TABLE` token, comma-joined, in table order.
- Invariants: pure function of `BP_KIND_TABLE`'s current contents — adding/removing a table entry changes this function's output with no other code change required.

## Data Models

### Domain Model
This feature adds two concepts to the REPL's existing session-scoped state (`xce_sched_repl_state_s`, repl_internal.h) — not a new top-level entity: a "round collapse state" (Requirements 2, 3) and an independent "resource display preference" (`units_mode`, Requirement 6). It follows the same pattern prior REPL task revisions used to add `output_path`/`replay_complete` to that struct.

### Logical Data Model
New fields on `xce_sched_repl_state_s` (illustrative — exact types finalized at implementation time):

| Field | Type | Purpose |
|-------|------|---------|
| `collapse_active` | `bool` | Whether a collapsed range is currently open |
| `collapse_start_round` / `collapse_end_round` | `uint64_t` | Round-number bounds of the active range |
| `collapse_start_ts` / `collapse_end_ts` | `uint64_t` | Sim-time bounds, for the collapsed range's time span (Requirement 3.4) |
| `collapse_dispatch_count` / `collapse_preempt_count` | `int` | Cached fingerprint counts |
| `collapse_pend_jobids` / `collapse_n_pend_jobids` | pointer + count | Cached fingerprint pend-jobid set, in whatever list form the Jobid Range Compressor consumes |
| `units_mode` | enum (`raw`/`human`/`both`) | Requirement 6's resource-value display mode, settable via `set units`; defaults to `both` |

**Consistency & Integrity**: all fields reset together whenever a range flushes (print, then clear) or a new REPL session begins — `repl_state_ensure()`'s existing `calloc` zero-initializes new fields with no additional lifecycle code required, **except** `units_mode`: unless `both` is deliberately given `enum` value 0, `calloc`'s zero-initialization will not select it, so `repl_state_ensure()` must explicitly set it to `both` after allocation regardless of the enum's member order — the one field in this struct where zero-initialization alone cannot be relied upon to be the correct default.

No Physical Data Model, API Data Transfer, or Event Schema subsections apply: this feature has no persistent storage, no network wire format, and adds no new wire-level contract — it stays inside `scheduler-core`'s existing "simulation isolation" boundary (repl.h's own header comment: no `scheduler.h`, no `client.h`, no capnp types).

## Error Handling

### Error Strategy
Consistent with existing `repl.c`/`repl_break.c` conventions: allocation failures print an existing-style `"[sim] ..."` message and degrade to the safest available behavior rather than crashing or losing information, matching `repl_state_ensure()`'s own "never load-bearing, degrade to unattended" precedent.

### Error Categories and Responses
- **Allocation failure while building a round's fingerprint or growing the collapsed-range pend-jobid buffer**: fall back to printing that round's line directly (uncollapsed) rather than losing the round's information — preserves Requirement 2.1's obligation to report every round with decisions.
- **`repl_fmt_jobid_range()` buffer too small**: returns a negative value; the caller falls back to the existing uncompressed per-id listing for that call — never truncates silently, never crashes, preserves the "never drop a jobid" invariant even in a buffer-constrained edge case.

### Monitoring
No new logging framework is introduced — the existing `[sim] ...` stdout convention remains the only user-facing signal; this is an interactive CLI tool, not a daemon, so no zap/WARN-level logging applies here.

## Testing Strategy

### Unit Tests (new `test_repl_output_compression.c`)
- Contiguous indices from one array base compress to a single `[lo-hi]` range (Requirement 4.1).
- Non-contiguous indices compress to comma-joined sub-ranges within one base (Requirement 4.2).
- Multiple array bases in one listing compress independently (Requirement 4.3).
- A mixed array + non-array listing shows non-array ids individually and compresses only the array ids (Requirement 4.4).
- A buffer too small to hold the full listing returns a negative value and writes nothing past `bufsz` (Error Handling).

### Unit Tests (new `test_repl_resource_formatting.c`)
- `repl_is_bytes_valued_resource()` returns true for exactly `mem`, `tmp`, `maxmem`, `maxtmp`, `maxswap`, and false for `swp`, unitless resources (e.g. `r1m`), and an unknown/custom name (Requirement 6.1).
- `repl_format_bytes_human()` adaptively scales across the full B/KB/MB/GB/TB unit range, including exact boundary values (e.g. exactly 1024 bytes renders as `1.00 KB`, not `1024.00 B`) (Requirement 6.2).
- `swp`'s value is converted MB→bytes before formatting, so a raw `swp` value renders on the same adaptive scale as a byte-native resource of equivalent magnitude (Requirement 6.1).
- A whole-number unitless value (e.g. `22.0`) renders with no trailing `.00`; a fractional one (e.g. `0.57`) still renders with its decimal places (Requirement 6.5).

### Integration Tests (extend `test_repl_breakpoints_inspection.c`, `test_repl_snapshot_params.c`)
- `set units raw`/`set units human`/`set units both` each select the corresponding rendering for a subsequent `info worker`/`info resource` call; an unrecognized `set units` value is rejected with the existing `set: unrecognized parameter` error shape (Requirement 6.2, 6.3).
- `info worker`'s resource listing renders as a grouped, aligned table with `used=` shown only for consumable/capacity resources, matching the hand-validated shape against a real captured worker's resource dump (Requirement 6.6).
- `info worker`'s assigned-job listing, given a job array's worth of contiguous indices, renders as a single compressed range via the same helper Requirement 4 already covers (Requirement 4.5).
- `break`'s unrecognized-event-type error message lists the complete, current `BP_KIND_TABLE` token set (Requirement 7.5).

### Integration Tests (extend `test_repl_execution_control.c`, child-process output capture)
- `continue` across multiple idle rounds (zero decisions) produces one collapsed range line, not one line per round (Requirement 3.1).
- `continue` across rounds sharing an identical nonzero pend set produces one collapsed "unchanged" range line — the exact shape hand-validated against the real 108-round replay that motivated this spec (Requirement 3.2).
- A structural event mid-`continue` ends any active collapsed range and prints its own line before continuing (Requirement 1.3, 3.5).
- A breakpoint hit mid-collapse flushes the active range before entering the command loop (Requirement 3.6).
- Replay completion (EOF) flushes any active collapsed range before the existing "replay complete" banner (Requirement 3.6).
- `step`-mode output is unchanged versus this feature's pre-implementation baseline — regression check alongside the existing `test_should_pause_fires_on_exactly_the_nth_session`-style fixtures (Requirement 5.1, 5.3).
- A job's tracked status changing between rounds is reflected in the next round summary even under a fixture that never emits a synthetic `DISPATCH` decision — exercises the state-derived-pend-set design decision independent of bug-1611's resolution (Requirement 2.4).
- The REPL prompt is the plain `"(sim) "` before any session begins, then includes the current session count and a human-readable timestamp once at least one session has run, in both `step` and `continue` modes (Requirement 5.5).
- `help`'s output groups commands into the four labeled sections, uses one consistent column width throughout, lists `info config` alongside the other `info` subcommands, and includes the complete `break` event-type list (Requirement 7.1-7.4).

### E2E
Not applicable beyond the integration tests above — this feature has no HTTP/API surface; the child-process harness already exercises the real `xce_scheduler` binary end-to-end at the process level.

## Optional Sections
Security Considerations, Performance & Scalability, and Migration Strategy are omitted: this is a local, single-operator interactive CLI output change with no new attack surface, no persistent or high-load data path (round counts are bounded by replay length, already tolerated by existing `step` output), and no schema or data migration.

## Supporting References
None beyond `research.md` — no external API, vendor library, or long schema requires separate documentation.