fix(compaction): render APPROX_ATTRIBUTION kept_tail/pre ratio truthfully (>100% basis artifact)

[Kyzcreig]

What

The in-turn A-floor observability marker printed kept_tail 75378 / pre 74447 = 101.3% in #logs (2026-07-01) — a confusing >100% that reads as "kept more than existed."

Why it's not a real bug (just a display artifact)

raw_tail_tokens is a documented upper bound estimated on the raw (pre-sanitize) suffix; pre_tokens is counted on the sanitized/in-context basis. Different bases → the raw bound can legitimately exceed pre. Compaction itself is unaffected — only the displayed fraction was misleading.

Fix (display-only)

New _fmt_gross_frac() renders the ratio truthfully:

• raw > pre → ≥100% (raw-tail bound N ≥ pre M; raw vs sanitized basis) (bounded + basis note)
• within bound → plain percent (unchanged)
• pre=0 → n/a (pre=0) (no divide-by-zero)

The threshold comparison is unchanged — a bound over the 10% gross-error line still correctly triggers the two-line degrade. Only what's printed changed.

Tests

3 new unit tests incl. the real captured 75378/74447 case; 116 compaction tests green, ruff clean.

Reported via Apollo's 5-cron-alert investigation (Phase 6). Spec: ~/.hermes/plans/2026-07-01_five-cron-alerts-FIX-PRD.md.

[greptile-apps]

Greptile Summary

Adds _fmt_gross_frac(), a small display-only helper that formats the raw_tail_tokens / pre_tokens ratio truthfully when the raw upper bound exceeds the sanitized pre-total. The compaction threshold comparison and all downstream logic are untouched.

• _fmt_gross_frac handles three cases: pre=0 → "n/a (pre=0)"; frac > 1.0 → bounded "≥100%" with a basis note; otherwise plain "{:.1%}". Three new unit tests cover the exact live-captured 75378/74447 case, the within-bound path, and the zero-denominator guard.

Confidence Score: 5/5

Display-only change with no effect on compaction logic or threshold comparisons; safe to merge.

The change is surgically scoped to two log-format call sites. The threshold check that drives real behavior (_gross_frac > _APPROX_GROSS_MAX_FRAC) is untouched. The new helper correctly handles all three reachable states (over-bound, within-bound, zero denominator), and the live-captured regression case is explicitly asserted in the test suite.

No files require special attention.

Important Files Changed

Filename	Overview
agent/conversation_compression.py	New _fmt_gross_frac helper added; two call sites swapped from raw :.1% to the helper. Threshold comparison (_gross_frac > _APPROX_GROSS_MAX_FRAC) is untouched and continues to use the raw float.
tests/agent/test_compaction_stats_degrade_observability.py	Three new unit tests covering the >100% bound case (real captured values), within-bound plain-percent path, and zero-denominator safety. All assertions are tight and specific.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["approx_attribution path\n_gross_tok, _pre_tok computed"] --> B{"_gross_frac > 10%?\n(threshold check — unchanged)"}
    B -- "yes" --> C["_warn_compaction_stats_once\n'degraded … two-line'\n_inturn_stats = None"]
    B -- "no" --> D["_warn_compaction_stats_once\n'APPROX_ATTRIBUTION in-turn …'\n_inturn_stats = _cand"]
    C --> E["_fmt_gross_frac(_gross_tok, _pre_tok)"]
    D --> F["_fmt_gross_frac(_gross_tok, _pre_tok)"]
    E --> G{"pre_tok <= 0?"}
    F --> G
    G -- "yes" --> H["return 'n/a (pre=0)'"]
    G -- "no" --> I{"gross/pre > 1.0?"}
    I -- "yes" --> J["return '≥100% (raw-tail bound N ≥ pre M; raw vs sanitized basis)'"]
    I -- "no" --> K["return '{:.1%}'"]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["approx_attribution path\n_gross_tok, _pre_tok computed"] --> B{"_gross_frac > 10%?\n(threshold check — unchanged)"}
    B -- "yes" --> C["_warn_compaction_stats_once\n'degraded … two-line'\n_inturn_stats = None"]
    B -- "no" --> D["_warn_compaction_stats_once\n'APPROX_ATTRIBUTION in-turn …'\n_inturn_stats = _cand"]
    C --> E["_fmt_gross_frac(_gross_tok, _pre_tok)"]
    D --> F["_fmt_gross_frac(_gross_tok, _pre_tok)"]
    E --> G{"pre_tok <= 0?"}
    F --> G
    G -- "yes" --> H["return 'n/a (pre=0)'"]
    G -- "no" --> I{"gross/pre > 1.0?"}
    I -- "yes" --> J["return '≥100% (raw-tail bound N ≥ pre M; raw vs sanitized basis)'"]
    I -- "no" --> K["return '{:.1%}'"]

Loading

_{Reviews (1): Last reviewed commit: "fix(compaction): render APPROX_ATTRIBUTI..." | Re-trigger Greptile}