Add the Article 15 resilience axis to the EU AI Act tags

[Igarez]

Follow-up to the discussion in NVIDIA#2094, as a patch rather than a comment: @leondz asked for the clarifications to go into the docs and, in short form, into the tag descriptions, and @erickgalinkin said the tags should extend further. It targets ai-act-mapping so it can be reviewed and squashed into that PR, or dropped if the reading is wrong.

Rationale and the Article-by-Article sourcing are in my comment on #2094. In short: the ten euai: tags describe what a model produces or is used for — the Article 5 and Article 55 axis, where 26/189 is about right. Article 15(5) is the other axis, what a high-risk system must withstand, and it names the categories itself: data poisoning, model poisoning, model evasion, confidentiality attacks and model flaws.

What changes

• Five euai:robustness:* tags added, sourced to Art. 15(5). dan, tap, goat, encoding, suffix, promptinject, latentinjection, web_injection, leakreplay, divergence, sysprompt_extraction and glitch become Article 15 evidence instead of no evidence at all.
• euai:deception:hallucination → euai:accuracy:hallucination, sourced to Art. 15(1) accuracy. Hallucinating is not deceiving, which was the objection; accuracy is what those probes measure.
• euai:harmful:sexualminor dropped. Not an Article 5 prohibited practice: its basis is Directive 2011/93/EU, and the DSA where a platform deploys the system.
• Every description names its provision, so a garak report can be cited in an Annex IV technical file without a translation step. protected_class is sourced to Art. 10(2)(f) as well as 5(1)(g), which is what reaches lmrc.Deadnaming — Case C-13/94 P v S holds that discrimination on gender reassignment is discrimination on grounds of sex.
• Docs: the two axes, and three coverage limits stated outright — thin data/model poisoning coverage, no minors tag and why, and Article 10/53 duties being outside what a black-box probe can evidence.

What this deliberately does not do

adaptive_attacks, dra, sata, fitd, phrasing, audio, doctor, continuation and topic are left unassigned rather than guessed at from the repository alone; most look like euai:robustness:evasion. goodside.Davidjl is likewise left alone. euai:robustness:data_poisoning has no probe today and the docs say so.

Checks

Every euai: tag in the tree resolves against tags.misp.tsv (no dangling references after the rename and the drop); black output is unchanged from the base branch; the reStructuredText added parses clean under docutils, with the same single INFO the section already produces on ai-act-mapping.

[erickgalinkin]

I am very, very grateful for your additions here. I made some updates independently and will close this PR. There are a few places I don't fully align with your analysis but please feel free to continue discussing with me on the main thread.