feat(core-engine): per-file dynamism replaces the flat language irc in four risk equations (#2719)

[squid-protocol]

Child of #2716; stacked on #2721 (PR 1). After PR 1 the per-language irc was still a flat integer summed into six formulas in six different units. This PR gives each of those formulas the term it can justify, one commit per formula.

Per formula

formula	commit	change	why
documentation	46890c5d	irc → per-file dynamism (reflection_metaprogramming, tuning documentation.dynamism_weight 1.0)	undocumented runtime-decided behaviour is exactly what needs documenting; the formula read neither signal before
cognitive load	3263483d	heat_density reads dynamism at its existing 5× weight; the irc / mass_loc pseudo-hit is removed	reflection and eval are the same opacity to a reader; one definition. Counting eval here and in safety is two risk dimensions, not a double count
concurrency	266dff79	language term removed (irc × 0.1, tuning key deleted)	no strictness column is about concurrency and dynamism is not concurrency; the inputs are the file's own spawns and locks
state flux	d4d17b07	language term removed (irc × 0.15, tuning key deleted)	"implicit mutability defaults" is not a strictness column and dynamism is not mutation
safety	—	keeps the strictness irc	the language-hazard claim (C's memory model) is defensible in a safety score; high_risk_execution is already the ×4 attack term
tech debt	—	keeps the strictness irc × 0.5	"things left unsaid" is maintenance overhead; dynamism is not debt

SignalProcessor._dynamism() is the single definition: the file's reflection_metaprogramming count. The first cut also summed high_risk_execution; attribution showed that signal is the safety attack vocabulary across the registry — panic!/todo!, os.Exit/log.Fatal, System.exit, STOP RUN, rm -rf/sudo — carrying eval/exec only in the dynamic languages, and it was importing aborts and destructive commands into cognitive load (three STOP RUNs in a 31-line COBOL file: +33). It is read at ×4 by safety, where it belongs; python's eval/exec therefore do not count as dynamism until a dedicated signal exists. Pointer arithmetic, macros and type bypasses stay in their own equations too.

Net: the language-level irc is read by exactly two formulas, each with a written reason a reviewer can disagree with; four formulas lost a constant that corresponded to nothing measurable. keyword-rosetta's tier_sensitive set shrinks to those two rows on its own (it AST-reads which formulas take irc).

Tests and tools

• test_uef_length_invariance.py: the 51-LOC pins are re-pinned per commit for exactly the equation that commit changed (the re-pin script refuses to run if any other equation moved); the engine PR 1: language-strictness table + measured fidelity coefficient replace _get_tier (child of #2716) #2718 exemption block is gone. The strictness-direction test still passes — equations that no longer read irc are equal across profiles.
• audit_risk_equations.py: _calc_concurrency and _calc_state_flux move to NOT_TIER_PARAMETERIZED (the tool asserts on this). audit_length_invariance.py updated. Both run clean.
• Dead tuning key documentation.irc_mult (never read) replaced by dynamism_weight; concurrency.irc_mult and state_flux.irc_mult removed.

What moved

Golden master regenerated in both venvs and attributed with golden_diff.deep_compare against PR 1's fixture: 5,231 substantive diffs, 0 topological, confined to the four changed columns and their aggregates — Safety, Tech Debt, Testing and everything structural are untouched, as intended.

column	cells	mean Δ	range	reading
Cognitive Load	1,654	−1.24	−7.47 … −0.01	only down: the irc / mass_loc pseudo-hit is gone and reflection was already in the heat term
State Flux	714	−1.73	−5.98 … −0.01	only down: term removed
Concurrency	291	−2.65	−7.89 … −0.01	only down: term removed
Documentation	643	−0.41	−25.84 … +87.89	down where the language irc exceeded the file's reflection count (jcl −4.7, livecode −4.3, assembly −2.9 means); up where the file is reflection-heavy

The upward Documentation tail is shell: 15 of the crucible's 46 shell files jump ≥ 25 points because the shell reflection_metaprogramming rule counts every $(...), backtick and plain ${var} (configure-helper.sh: 1,019 hits in 2,823 lines) — filed as #2722. That is the per-file term doing its job: it made a vocabulary over-match visible that the flat language constant had hidden. The fix belongs to the rule, not the formula.

The first cut of this PR summed high_risk_execution into dynamism; its attribution showed Cognitive Load moving up by +33 on a 31-line COBOL file with three STOP RUNs. That is what led to narrowing the definition (bec8522b).

Gauntlet

• Full suite: 7,566 passed, 3 skipped, 9 xfailed (the 51-LOC pins re-pinned per commit, see above).
• audit_check.py: all clear at every commit (ruff baseline regenerated for line-shifts; mypy 2-error baseline; dead-key; ast-accuracy).
• tree_sitter_accuracy_audit.py --ci --all: 30/30. tri_comparison_chart.py --all --ci: 3/3.
• rosetta_audit.py: 46/46, 0 regressions — §2 signals untouched, so no corpus re-bless owed; the risk_* rows move on the next bias-history regen and tier_sensitive shrinks to safety + tech debt on its own.
• audit_risk_equations.py and audit_length_invariance.py: no direction or invariance flags.
• Golden master: blessed in both venvs (crucible_check.py --update --yes), scoped diff above.

Closes #2719

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab