Drop the linkermap Metrics Pipeline

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Goal: Remove the legacy linkermap-based code-size pipeline (CMake targets, tools/metrics.py, the tools/linkermap dependency, and --engine linkermap) once the membrowse engine has proven itself in daily CI use, leaving membrowse as the only size-diff engine.

Architecture: No new architecture — this is a subtraction. family_add_linkermap (CMake), tools/metrics.py, the tools/linkermap entry in tools/get_deps.py, and the --engine/--combined linkermap code paths in tools/metrics_compare_base.py all go away; family_add_bloaty and family_add_membrowse are untouched (bloaty is a separate, independent analysis tool with its own target).

Tech Stack: CMake (hw/bsp/family_support.cmake), Python 3.13 stdlib (tools/metrics_compare_base.py, tools/get_deps.py).

Spec: none — split out of the rework-metrics branch (docs/superpowers/specs/2026-08-25-rework-metrics-design.md), which made membrowse the default engine but kept linkermap alive as the only engine --combined/--ci can use.

Origin: split out of the rework-metrics branch's final-review fix wave. Delete this file when its own PR lands.

What is already established

  • membrowse is already the default engine. Commit 4d9629354 (“ci: membrowse-only size analytics - drop linkermap metrics pipeline, gate uploads on pinned boards”) switched CI uploads to membrowse; tools/metrics_compare_base.py --engine defaults to membrowse (see its --engine argument).
  • Per-file equivalence between the two engines is proven, not assumed. During this branch's build, an object-file-keying bug (membrowse's source_file truncates to a bare basename, so an absolute-path filter matched nothing) was found and fixed by keying/filtering on object_file instead (tools/membrowse_compare.py module docstring + per_file_sizes). After the fix, an equivalence re-run passed: tusb.c attributed +40 B on both engines (.superpowers/sdd/2026-08-25-rework-metrics/progress.md:45, commits 314520011..39c53cdd4). That is the evidence a wholesale engine swap is safe for the per-file delta table, not just for totals.
  • CI no longer asks linkermap to run. tools/build.py's cmake configure step used to pass -DLINKERMAP_OPTION=-q -f tinyusb/src on every build (removed in this same fix wave, see F7 of the final-review report) and no CI build target requests <example>-linkermap or examples-linkermap any more — those targets are dead weight in a normal CI run, built only when tools/metrics_compare_base.py --engine linkermap explicitly asks for them.
  • The remaining necessary use of --engine linkermap is narrow and identified. It is --combined/--ci only: metrics_compare_base.py hard-refuses --combined with the membrowse engine (parser.error('--combined is not yet supported with --engine membrowse ...')), and --ci implies --combined by setting args.combined = True. Every other invocation (single example, all examples on one board) already defaults to membrowse. .claude/skills/code-size/SKILL.md and .claude/skills/membrowse/SKILL.md both point users at --engine linkermap for exactly this one case.
  • Inventory of what removal touches (grep -rl linkermap across the tree, minus this branch's own planning docs):
    • hw/bsp/family_support.cmake — family_add_linkermap(), the <target>-linkermap / examples-linkermap CMake targets, and its call site beside family_add_bloaty().
    • tools/metrics.py — the whole file (report generation + compare subcommand); nothing else imports it once the linkermap branch of metrics_compare_base.py is gone.
    • tools/get_deps.py — the tools/linkermap entry in deps_all.
    • tools/metrics_compare_base.py — the --engine flag (collapses to membrowse-only), the want_linkermap branch in the build loop, generate_metrics(), and the combined-mode linkermap aggregation block.
    • test/hil/test/test_metrics_compare_base.py — whatever linkermap-specific cases exist there today need triage (drop, or repoint at membrowse if they test shared plumbing).
    • CLAUDE.md, .claude/skills/code-size/SKILL.md, .claude/skills/membrowse/SKILL.md — every --engine linkermap / tools/linkermap mention.

What remains (not started)

  1. Land pr3887-membrowse-combined.md first (the blocker below).
  2. Delete family_add_linkermap and its two custom targets from hw/bsp/family_support.cmake; drop the call to it inside the block that currently calls family_add_bloaty(${TARGET}) / family_add_linkermap(${TARGET}) / family_add_membrowse(${TARGET}).
  3. Delete tools/metrics.py.
  4. Delete the tools/linkermap entry from deps_all in tools/get_deps.py; update CLAUDE.md's worktree-symlink note (it lists tools/linkermap as an example deps_all key — pick a different example or drop the parenthetical).
  5. Collapse tools/metrics_compare_base.py to membrowse-only: drop --engine, the want_linkermap branch, generate_metrics(), and the combined-mode linkermap aggregation — --combined/--ci then call the membrowse combine path added by the other follow-up instead of erroring.
  6. Update test/hil/test/test_metrics_compare_base.py, .claude/skills/code-size/SKILL.md, .claude/skills/membrowse/SKILL.md to match.
  7. Full sweep validation: pre-commit run --all-files, plus one real --ci run to confirm the membrowse combined path (once it exists) produces the same shape of report the linkermap combined path did.

Why this is a separate PR

  • It is blocked, not merely deferred: --combined/--ci has no membrowse implementation yet (pr3887-membrowse-combined.md), and linkermap is the only engine that path can use today. Removing linkermap before that lands would delete the only way to run a full CI-family size sweep.
  • “Proven in daily use” is a time-based gate, not a code-based one. The rework-metrics branch is membrowse's first real deployment; the right trigger for this removal is “membrowse has been uploading real CI numbers without a regression for some weeks,” which cannot be established inside the PR that just switched the default.
  • Bundling it into this PR would blow up an otherwise-reviewable diff. This branch's scope is “rework metrics to run on membrowse”; deleting a second engine wholesale, plus its CMake targets, plus a getdeps entry, plus a doc sweep, is a distinct, independently reviewable change with its own regression risk (anyone who has --engine linkermap in a local script or bookmark loses it).