fix(context_breakdown): search the volatile tier for the skills index

[pierrenode]

Summary

#37117 (salvaged and merged today as #77696) moved the <available_skills> block from the stable band of the system prompt to the volatile band, so skill edits don't invalidate the cached identity prefix. A same-day CI-caught follow-up fixed hermes_cli/prompt_size.py's compute_prompt_breakdown() for this move.

agent/context_breakdown.py — which powers the /context and /context all commands (shipped via #72242) — has its own independent copy of the same _SKILLS_BLOCK_RE regex, used in two places (compute_session_context_breakdown(), compute_context_details()), and both were missed by the follow-up fix: they still search only the stable tier.

Empirically confirmed before writing the fix: with a skill loaded, /context's "Skills" category is silently absent from the breakdown, and /context all lists zero skills.

Fix

• Both functions now search the volatile tier first, falling back to stable for sessions whose cached prompt predates fix(system_prompt): move the skills index out of the stable band to preserve the cached prefix #37117 (mirrors prompt_size.py's fix).
• compute_session_context_breakdown() also now strips the found skills block out of the volatile tail before folding it into the system_prompt text, so skills tokens are attributed only to their own category rather than double-counted once they're found (a bug that would otherwise have been introduced by fixing the lookup alone).

Test plan

• Updated test_breakdown_includes_major_categories's fixture to place <available_skills> in the volatile tier (matching current production reality) rather than stable — this is why the bug wasn't caught: the existing test's mock reflected the pre-fix(system_prompt): move the skills index out of the stable band to preserve the cached prefix #37117 layout.
• Added test_skills_index_falls_back_to_stable_for_legacy_sessions for the legacy fallback path.
• Added test_skills_index_not_double_counted_in_system_prompt proving the volatile-tail stripping fix.
• Added test_context_details_finds_skills_in_volatile_tier — compute_context_details() had no prior direct behavioral test coverage at all.
• Mutation-verified: reverting the fix makes all three new/updated assertions fail with the exact symptom described above.
• Full neighboring test sweep (context breakdown, prompt_size, system_prompt, status command — 61 tests) passes.
• ruff check clean on all changed files.

[86523553]

Independent reproduction on 0.21.5+2446.g9fc7f17906 (Windows 11, Python 3.14.7, Desktop app, platform=cli toolset resolution, 272 installed skills) — confirms the diagnosis in this PR and adds the user-visible numbers.

from hermes_cli.prompt_size import _build_inspection_agent
from agent.system_prompt import build_system_prompt_parts
from agent.context_breakdown import compute_session_context_breakdown

a = _build_inspection_agent("cli")
p = build_system_prompt_parts(a)
print(len(p["stable"]), len(p["volatile"]))                                      # 9942 29260
print("<available_skills>" in p["stable"], "<available_skills>" in p["volatile"]) # False True
print([(c["id"], c["tokens"]) for c in compute_session_context_breakdown(a, [])["categories"]])

Output:

9942 29260
False True
[('system_prompt', 12229), ('tool_definitions', 17477), ('subagent_definitions', 1165)]

So on a fresh session: there is no skills category at all, and system_prompt is inflated to 12,229 tokens while the stable instruction tier is only 9,942 chars (~2.5K tokens) — the missing ~9.7K is the <available_skills> index living in the volatile tier. hermes prompt-size independently reports the skills index at 31,336 B for the same config, and it never surfaces as its own category.

Two user-visible symptoms in the Desktop Context usage popover, both from this root cause:

1. A Skills row never appears (zero-token categories are filtered out of categories in compute_session_context_breakdown), so the panel looks like skills are free.
2. System prompt absorbs the entire skills index, so it reads ~12K instead of ~2.5K — i.e. the single biggest fixed cost in the prompt is mislabeled as instructions.

Same payload feeds the Desktop popover, /context, and the gateway slash-command renderer, all of which inherit the mislabel (and /context all lists zero skills, as noted in the PR). +1 to landing this.